Machine learning classification of encrypted network traffic

[0001] The present application claims the benefit of Canadian patent application 3,197,148 filed on Apr. 16, 2023 entitled “Method and System for Classifying Encrypted Network Traffic Using Artificial Intelligence”, the entire contents of which are incorporated herein by reference.

[0002] The present invention relates to labeling and classification of network traffic, and in particular to methods and systems for labeling modeling and classification of the encrypted network traffic using artificial intelligence techniques.

[0003] A significant portion of network traffic is encrypted. Many applications tunnel traffic over HTTPS to provide data privacy to users. The introduction of Transport Layer Security 1.3 (TLS 1.3), rendered conventional methods of inspecting network traffic unsuitable to use for distinguishing legitimate traffic from illegitimate traffic, blocking illegal applications, enforcing policies, or detecting cyber threats hiding in encrypted traffic.

[0004] In addition, the proliferation of the encrypted traffic renders useless many of existing tool-chains for network situational awareness, network planning, and cyberspace operations, which are becoming unable to perform network forensics, obtain visibility into encrypted traffic, secure the network, differentiate subscriber offerings, ensure suitable Quality of Service; or conduct network planning.

[0005] Therefore, there is a need for developing improved methods and systems for labeling, modeling and classifying encrypted network traffic, which would solve or at least partially eliminate the above noted problems.

[0006] There is an object of the present invention to provide methods and systems for improved labeling, modeling and classification of the encrypted network traffic.

[0007] In accordance to one aspect, the invention provides a method of labeling network flows, implemented at a network appliance employing a hardware processor. The method comprises processes of:

• • [0008] acquiring a first set of records of training flow data containing unlabeled flows;
  • [0009] filtering the first set of records to produce a filtered set excluding records determined to be outliers according to respective criteria;
  • [0010] initializing a second set of labeled-flow records as an empty set; and
  • [0011] recurrently labeling successive samples of the filtered set. for at most a prescribed number of cycles.

[0012] Each sample is selected to not exceed a preselected number of records from the filtered set. Subject to a determination that the second set contains a number of records at least equal to a prescribed lower bound a semi-supervised learning model is trained based on the second set and used to attempt to label each record of the filtered set to produce a respective label suggestion. The label of each record of the sample is based on respective available label suggestions. The second set of labeled-flow records is updated to include successfully labeled records and successfully labeled records of the sample are removed from the filtered set.

[0013] The process of acquiring the first set of records comprises employing known models to produce, for each acquired unlabeled flow, at least one of: a DPI-suggested label; a suggested label according to a pre-trained machine-learning model; a suggested label according to a pre-trained deep-learning model; a rule-based suggested label; and a suggested label based on traffic-flow pattern.

[0014] The recurrent labeling process is terminated upon reaching any of the following conditions:

• • [0015] (a) a number of remaining unlabeled flows of the filtered set is less than a respective predetermined lower bound;
  • [0016] (b) a proportion of unlabeled flows of the filtered set that has been successfully labeled exceeds a respective predetermined upper bound; and
  • [0017] (c) a succession of a predetermined number of samples failed to label any unlabeled flow.

[0018] Selecting a sample is based on unbiased sampling of unlabeled records of the filtered set.

[0019] The method optionally comprises extracting from the filtered set a respective representative set of unlabeled flows of a reduced size applying one of two approaches. A first approach is unbiased selection of the unlabeled flows of the filtered set to pick a desired number of extracted unlabeled flows. The second approach is to apply k-means clustering of the unlabeled flows of the filtered set of unlabeled flows using a number of initial centroid seeds equal to a desired number of extracted unlabeled flow, where each record of the filtered set of records is presented as a respective vector of a multi-dimensional space.

[0020] The method further comprises presenting each record of the filtered set of records as a respective vector of a multi-dimensional space and clustering vectors of the filtered set to produce a requisite number of clusters not exceeding a network-appliance-specific upper bound. Suggesting a label for a flow may be based on examining the clusters.

[0021] The clustering process is preferably performed using the Density-Based Spatial clustering of Applications with Noise (DBSCAN) with a hypersphere radius and a minimum number of vectors within the hypersphere selected iteratively to yield the requisite number of clusters.

[0022] Each unlabeled flow comprises a stream of encrypted packets. Training may be performed as two-view co-training employing a first labeler based on features related to packet-size parameters and a second labeler based on time-domain packet-arrival patterns.

[0023] According to one embodiment, co-training is based on (a) supplying training data to the first labeler comprising cumulative labeled flow records determined at the first labeler in addition to switched labeled flow records of a respective most-recent sample determined at the second labeler, and (b) supplying training data to the second labeler comprising cumulative labeled flow records determined at the second labeler in addition to switched labeled flow records of a respective most-recent sample determined at the first labeler.

[0024] According to another embodiment, co-training is based on: (C) supplying training data to the first labeler comprising cumulative labeled flow records determined at the first labeler in addition to switched cumulative labeled flow records determined at the second labeler, and supplying training data to the second labeler comprising cumulative labeled flow records determined at the second labeler in addition to switched cumulative labeled flow records determined at the first labeler.

[0025] In accordance with another aspect, the invention provides a network appliance for classifying network flows according to a set of flow classifications comprising: (A) a labeler configured to receive a set of unlabeled training flows and label individual flows according to predefined flow classifications to produce a set of labeled training flows; (B) a modeler configured to determine classification of a flow comprising a stream of encrypted packets based on the set of labeled training flows; and (C) a classifier configured to classify, in real time, flows received from users of the network appliance using the modeler.

[0026] The labeler comprises: a pre-processing module; a sampling module configured to select successive samples of the filtered set, each sample not exceeding a preselected number of records from the filtered set; and a kernel module for labeling individual unlabeled flows of each sample.

[0027] The preprocessing module is configured to: filter the set of unlabeled training flows to produce a filtered set of unlabeled flows excluding outliers; cluster one of the filtered set and the representative set to produce a requisite number of clusters not exceeding a specified upper bound; and optionally extract a representative set of unlabeled flows of a reduced size.

[0028] The pre-processing module is further configured to employ known models to produce, for each unlabeled flow of the filtered set of unlabeled flows, at least one of: (I) a DPI-suggested label; (II) a suggested label according to a pre-trained machine-learning model; (III) a suggested label according to a pre-trained deep-learning model; (IV) a rule-based suggested label; and (V) a suggested label based on traffic-flow pattern.

[0029] According to a first embodiment, the network appliance is implemented as an integrated appliance comprising a training component, an operational component activated in real-time in a continuous operational phase, and a module for updating the training data based on outcome of the operational phase; the training component configured to produce models and corresponding parameters to be used for real-time classification.

[0030] According to a second embodiment, the network appliance as implemented as a partitioned appliance comprising a stand-alone training module communicatively coupled to a stand-alone operational module, each having a respective processor.

[0031] According to a third embodiment, the network appliance is implemented as a spatially distributed appliance comprising a central training module distributing training results to multiple operational modules and receiving results of real-time classifications from the operational modules.

[0032] In a preferred implementation, the labeler, the modeler, and the classifier are interconnected in a closed loop with the labeler receiving fresh training data from external sources and feedback data from the classifier comprising data relevant to real-time classified flow to be used for enhanced labeling.

[0033] The labeler, the modeler, and the classifier are configured to be activating concurrently to label received fresh training data, periodically re-synthesize the modeler, and continually classify flow data received from users directing classification results to respective users and simultaneously to the labeler through the closed loop.

[0034] A pruning filter may be used for trimming training data according to predefined criteria. The sampling module comprises a semi-supervised model for use in the kernel module. The semi-supervised model comprises a co-training module employing a first labeler based on packet-size parameters and a second labeler based on time-domain packet-arrival patterns.

[0035] Thus, improved methods and systems for labeling, modeling and classifying the encrypted network traffic using artificial intelligence techniques have been provided.

• • [0088] 100: Schematic of a network appliance (Trafficwiz™) for traffic-flow characterization implemented as an integrated appliance
  • [0089] 110: Training component of appliance 100 activated during a training phase
  • [0090] 112: A memory device storing Initial global training data
  • [0091] 114: A memory device holding updated training data
  • [0092] 116: AI model and parameters for use in real-time classification
  • [0093] 120: Operational component of appliance 100 activated in real-time in a continuous operational phase
  • [0094] 121: Traffic-flow classification requests received from users (also referenced as “clients”) of appliance 100
  • [0095] 122: Responses to classification requests 122
  • [0096] 123: Data relevant to classified users' traffic flows, defined in requests 121, supplied to process 114 for continually updating the training data
  • [0097] 130: Process of updating the training data based on output of the operational component
  • [0098] 150: Continuous recursive circulation during successive training epochs activating the training component 110, the operational component 120, and updating process 130 leading to a continually improved and updated training data 114
  • [0099] 200: A partitioned arrangement of network appliance 100
  • [0100] 220: A data network
  • [0101] 230: A stand-alone training module (a training part of integrated network appliance 100)
  • [0102] 240: A stand-alone operational module (a classification part of integrated network appliance 100)
  • [0103] 300: Hardware components of the appliance of FIG. 1 implemented as an integrated appliance
  • [0104] 320: A network interface of a training module of network appliance 100
  • [0105] 321: A dual channel to a data network
  • [0106] 325: A network interface of an operational module of network appliance 100
  • [0107] 326: A dual channel to a data network
  • [0108] 330: A memory device storing training software instructions
  • [0109] 335: A memory device storing operational software instructions
  • [0110] 340: A memory device storing current training data (initially storing pre-determined training data)
  • [0111] 342: A memory device storing records of training traffic flows that were not labelled during a training epoch for use in further analyses
  • [0112] 345: A memory device storing additional traffic-flow records acquired during the operational phase (real-time classification phase) of a training epoch to be used in a subsequent training epoch
  • [0113] 350: A processor; generally, an assembly of hardware processing units operated in parallel and/or in a pipelined arrangement
  • [0114] 400: Hardware implementation of the appliance of FIG. 2 in the form of two stand-alone modules, a training module and an operation module, that may be coupled directly or through a data network
  • [0115] 430: Implementation of stand-alone training module 230
  • [0116] 440: Implementation of stand-alone operational module 240
  • [0117] 500: A spatially distributed appliance for traffic-flow characterization comprising a training module and operational modules that may be spatially distributed
  • [0118] 510: Downstream training results to an operational module 550
  • [0119] 520: Upstream classification results from an operational module 550 to central training module 530
  • [0120] 530: A central training module similar to stand-alone training module 230
  • [0121] 532: A network interface of the central training module 530
  • [0122] 550: An operational module similar to stand-alone operational module 240
  • [0123] 552: A network interface of an operational module 550
  • [0124] 600: Integrated Trafficwiz™ appliance communicatively coupled to a plurality of clients through a data network as well as a distributed appliance comprising a stand-alone training module communicatively coupled to multiple stand-alone operational modules through the network, each operational stand-alone module being configured to communicate with the plurality of clients through the data network
  • [0125] 700: An overview of the main processes performed by a Trafficwiz™ appliance (integrated or partitioned)
  • [0126] 720: processes of integrated appliance 100 or partitioned appliance 200
  • [0127] 740: Training phase
  • [0128] 750: Real-time operational phase (classification of received traffic flows—from users or via the network)
  • [0129] 760: Labeler
  • [0130] 762: Salting module
  • [0131] 770: Modeler
  • [0132] 780: Classifier of fresh traffic flows received from clients
  • [0133] 800: Processes performed at integrated appliance 100 or partitioned appliance 200
  • [0134] 820: Process of acquiring a set of traffic records for training
  • [0135] 830: A process of acquiring or specifying a set of recognizable traffic-flow classifications
  • [0136] 822: A process of attempting to label each of the traffic records
  • [0137] 840: A process of synthesizing a traffic-flow modeler based on successfully labeled traffic-flow records
  • [0138] 850: A process of synthesizing a real-time traffic-flow classifier
  • [0139] 860: A process of continually receiving fresh traffic-flow data from the network or from users (clients) through network interface 325 or 552
  • [0140] 880: A process of real-time traffic-flow classifications using synthesized classifier generated by the modeler 850
  • [0141] 890: Classified traffic flows transmitted to concerned party
  • [0142] 900: An exemplary predefined set of homogeneous traffic flows, as well as a predefined set of composite traffic flows each formed as a multiplex of a respective subset of the set of homogeneous traffic flows
  • [0143] 1000: An exemplary training dataset of traffic flows to be labeled, according to the predefined homogeneous and composite traffic flows, undergoing preparatory (training) operations of filtering, clustering, and sampling
  • [0144] 1010: Unlabelled flows
  • [0145] 1012: Records of individual traffic flows
  • [0146] 1014: Records filtered out as outliers
  • [0147] 1020: Process of filtering of flow data
  • [0148] 1022: Records of individual retained traffic-flow records after filtering
  • [0149] 1030: Process of clustering and visualization of flow data
  • [0150] 1031: Records of cluster-1
  • [0151] 103j: Records of cluster-j,i≤j≤J, J being the total number of clusters
  • [0152] 1040: Process of sampling of training traffic-flow records, extracting subsets of a specified number of records
  • [0153] 1041: Records of subset-1 of training traffic-flow records
  • [0154] 104k: Records of subset-k of training traffic-flow records, 0≤k≤K, K being an index of the last subset to be drawn during a training epoch; K may be predefined or determined according to a criterion such as exhaustion of the training dataset of a current training epoch or labeling saturation.
  • [0155] 1100: Traffic-flow characterization
  • [0156] 1120: A specific traffic flow
  • [0157] 1121: Identifier of the specific traffic flow
  • [0158] 1122: Features (also referenced as “descriptors”, “metadata” or “statistics”) specific to the traffic flow
  • [0159] 1130: A tuple uniquely identifying the specific traffic flow
  • [0160] 1140: A table of a superset of predefined features characterizing traffic flows in general
  • [0161] 1150: A code indicating traffic-flow direction(s)—the terms “upstream” and “downstream” directions are synonymous with “forward” and “backward”
  • [0162] 1151: Values of code 1150: a code of “00” indicates that a corresponding feature is not relevant to specific traffic flow 1120, a code of “01” indicates that the specific flow is an upstream flow only, a code of “10” indicates that the specific flow is a downstream flow only, and a code of “11”, indicates a dual flow comprising an upstream flow and a downstream flow
  • [0163] 1160: Specifics of each feature of the superset of features
  • [0164] 1161: An index of a feature of the superset of features
  • [0165] 1200: Sorting (grouping) a superset of traffic-flow features according to predefined feature categories
  • [0166] 1210: A predefined set of Γ feature categories, Γ>1, individually the feature categories are referenced as 1212(1) to 1212(Γ)
  • [0167] 1220: A predefined superset, or an extracted set, of features
  • [0168] 1240: A process of sorting (grouping) the features according to predefined feature categories
  • [0169] 1260: Indices 1161 of features of each feature category (individually 1260-1 to 1260-M)
  • [0170] 1300: Optimal selection of a set of prominent features from a superset of predefined features
  • [0171] 1320: A process of feature data adaptation for processing in a genetic-algorithm module
  • [0172] 1330: A selected genetic-algorithm module
  • [0173] 1340: A produced set of prominent features; output of the genetic-algorithm module
  • [0174] 1400: A scheme for continuous training of the Trafficwiz™ appliance (whether integrated or distributed)
  • [0175] 1410: A process of continually acquiring global training traffic-flow records on a regular/periodic basis, such as training epoch of a predetermined duration (a day, a week, etc.)
  • [0176] 1420: A training stage for a current time-window (a current epoch)
  • [0177] 1430: A process of labeling samples (subsets) of the global training flow records
  • [0178] 1440: Processing of a selected subset
  • [0179] 1450: Processing of a last subset
  • [0180] 1460: Synthesis of the operational-modeler 770
  • [0181] 1500: Nesting of stages of scheme 1400, where each training phase of a continual sequence of training epochs includes multiple preprocessing stages, each preprocessing stage including multiple sampling stages, each sampling stage including multiple single-flow processing stages (referenced as Kernel stages)
  • [0182] 1510: Training-phase update performed during a training epoch
  • [0183] 1520: Pre-processing of a training dataset specific to the training epoch
  • [0184] 1530: Processing successive samples (successive subsets) of the training dataset
  • [0185] 1540: Sequentially processing individual traffic flows of each sample (each subset) of the training dataset
  • [0186] 1600: Recursive execution of the stages of scheme
  • [0187] 1610: Outer loop of processes executed during successive time windows (successive training epochs)
  • [0188] 1612: Refreshed training data for use in a current time window (a current training epoch)
  • [0189] 1614: Data of a current subset of training dataset
  • [0190] 1616: Data of a single traffic flow
  • [0191] 1620: Intermediate loop of processes executed during successive sampling periods, i.e., periods of attempting to label successive subsets of traffic flows
  • [0192] 1630: Processing of one traffic flow at a time until all flows in the subset are processed—exhaustive selection
  • [0193] 1700: A method of labeling a massive set of unlabeled or tentatively labeled traffic-flow records, based on progressively modifying the composition of the massive set as traffic-flow records get labeled
  • [0194] 1710: A memory device storing a superset of M unlabeled, or partially labeled, traffic-flow records. M being a relatively large integer
  • [0195] 1720: A process of extracting a sample of at most a predefined number, L, of traffic-flow records from the training dataset stored in memory 1710. L<<M
  • [0196] 1730: A process of determining whether all records from original M have been processed or that successive executions of process 1750 failed to label pending records
  • [0197] 1740: Loop of recursive processes executed subject to a need for continuity as determined in process 1760
  • [0198] 1750: A process of machine-learning or deep-learning limited to handle at most L records
  • [0199] 1760: Process of updating the superset of training records stored in memory 1710 to account of any successful labeling of process 1750
  • [0200] 1780: Process of storing data relevant to records successively labeled in process 1750
  • [0201] 1790: Process of terminating the labeling stage
  • [0202] 1800: A flow chart of a specific implementation of method 1700
  • [0203] 1810: Process of initializing a file, referenced as a “label file”, for holding labeled records of traffic flows. Initially, the labeled file is set as an empty file
  • [0204] 1820: Process of assembling records of training data in a “global file”—for a first training epoch, the global file contains only unlabeled records, for subsequent training epochs, the global file contains a mixture of prior residue records as well as newly acquired unlabeled records
  • [0205] 1835: Loop of processes 1830, 1840, 1845, 1850, 1860, 1865, 1870, 1880
  • [0206] 1840: Process of selecting a sample (a subset) of a preselected number of records from the global file
  • [0207] 1845: Checking count of labeled flows to determine whether use of a semi-supervised learning model is appropriate
  • [0208] 1850: Process of training a semi-supervised learning model based on labeled records
  • [0209] 1860: Processes of attempting to label each record of the sample
  • [0210] 1870: Process of inserting successfully labeled records into the label file
  • [0211] 1880: Process of removing successfully labeled records of the sample from the global file and revisiting branching point 1830
  • [0212] 1900: An example of labeling of successively selected subsets of the training dataset
  • [0213] 1901: Count of initial unlabeled records
  • [0214] 1902: Count of initial labeled records (zero)
  • [0215] 1920: Counts of unlabeled records after processing successive samples
  • [0216] 1940: Counts of labeled records after processing successive samples
  • [0217] 1950: Counts of all records of training data after processing successive samples
  • [0218] 1960: Count of residual records of the global file that were not labeled
  • [0219] 2000: Labeling attainment for successive subsets of the training dataset based on changing proportion of labeled traffic-flows records of the global traffic-flow records
  • [0220] 2010: Successive samples; individually 2010(j), j≥0 till a last sample of a current epoch
  • [0221] 2020: Labeled and unlabeled records of each processed sample with updates of the superset of records
  • [0222] 2100: An exemplary global set of traffic flows each traffic flow presented as a vector of a multi-dimensional space, a simplified two-dimensional space is used for ease of presentation
  • [0223] 2120: A record of a traffic flow; a vector of flow characteristics
  • [0224] 2140: A hypersphere of a radius ρ containing a number, ν, of records; two hyperspheres, individually 2140(1) and 2140(2), are indicated of parameters (ρ₁, ν₁) and (ρ₂, ν₂).
  • [0225] 2200: Exemplary clusters of the global set of traffic flows
  • [0226] 2300: Successive samples (subsets of the training dataset) 2320 drawn from the global set of traffic flows
  • [0227] 2320: Individual samples individually identified as 2320(1) to 2320(8)
  • [0228] 2400: Details of sampling the global file with partial replacement
  • [0229] 2410: Exemplary clusters of the flows of FIG. 21, individually referenced as 2410(1) to 2410(5)
  • [0230] 2420: Successively drawn samples, S₁ to S₈, from the global file, with partial replacement, individually referenced as 2420(1) to 2420(8)
  • [0231] 2500: Processes of the pre-processing stage of FIG. 15
  • [0232] 2510: Process of acquiring raw unlabeled flow data
  • [0233] 2512: A memory device holding unlabeled flow data
  • [0234] 2520: Process of filtering the raw unlabeled flow data to exclude outliers
  • [0235] 2522: A memory device holding filtered raw unlabeled flow data; optionally, the filtered data may be stored in a division of memory device 2512
  • [0236] 2530: Process of clustering the filtered flow data
  • [0237] 2532: A memory device storing clustering data
  • [0238] 2534: Process of generating a flow-scatter plot and a packet-sequence plot
  • [0239] 2550: Process of producing a DPI suggested label for each unlabelled flow
  • [0240] 2555: A process of producing a suggested label for each unlabelled flow based on a pre-trained (apriori-trained) ML/DL model
  • [0241] 2560: A process of producing a suggested label for each unlabelled flow based on specific rules
  • [0242] 2580: A process of generating augmented filtered raw unlabeled flow data indicating the suggested labels produced in processes 2550, 2555, 2560 and 2565
  • [0243] 2585: A memory device storing the augmented filtered raw unlabeled flow data
  • [0244] 2600: Processes of the sampling stage of FIG. 15
  • [0245] 2620: A branching point at which continuity or completion of the sampling stage of a current training epoch is determined
  • [0246] 2610: A process of initializing a set of labeled flows as an empty set
  • [0247] 2640: A process of selecting a sample (a subset) of the augmented records for processing in the kernel stage
  • [0248] 2642: Module to train a semi-supervised model
  • [0249] 2650: Recursion of selecting samples of the augmented unlabeled flows until completion of the sampling stage for a current training epoch
  • [0250] 2670: A process of updating the set of labeled flows based on receiving results from the Kernel module
  • [0251] 2680: Indicate that processing of a current sample is complete
  • [0252] 2700: Processes of the kernel stage of FIG. 15
  • [0253] 2720: A branching point at which continuity of processing records of a current sample, or completion, is determined
  • [0254] 2725: Process of selection of an unlabeled flow of a current sample (current subset) of flows
  • [0255] 2730: A loop of individual flow processing
  • [0256] 2740: A branching point which initially bypasses process 2745
  • [0257] 2745: A process of suggesting a label using a semi-supervised model
  • [0258] 2750: A process of producing a system suggested label based on all label suggestions of the current flow
  • [0259] 2760: A process of verifying the system suggested label
  • [0260] 2765: A branching point at which a current flow remains unlabeled or proceed to a further process of machine evaluation
  • [0261] 2770: A process of machine evaluation of the current suggested labels
  • [0262] 2775: A branching point at which the current flow remains unlabeled or transferred from the global file to the label file
  • [0263] 2780: A process of updating the global file and the label file to account for successful labeling of the current flow
  • [0264] 2800: Processes of labeling subsets of the training dataset for a case of a relatively small global set of traffic flows and, consequently, a relatively small sample size
  • [0265] 2810: A sample space of 40 unlabeled flows
  • [0266] 2812: A first sample (first subset, 10 flows) of sample space 2810
  • [0267] 2814: A number (six) of successfully labeled flows of the first sample
  • [0268] 2820: An updated sample space (34 flows) comprising sample space 2810 excluding labeled flows of the first sample (six flows)
  • [0269] 2822: A second sample (second subset, 10 flows) drawn from the updated sample space 2820
  • [0270] 2824: A number (eight) of successfully labeled flows of the second sample
  • [0271] 2830: An updated sample space (26 flows) comprising sample space 2820 excluding labeled flows of the second sample (eight flows)
  • [0272] 2832: A third sample (third subset, 10 flows) drawn from the updated sample space 2830
  • [0273] 2834: A number (eight) of successfully labeled flows of the third sample
  • [0274] 2835: Remaining unlabeled flows (18 flows) after processing the third sample
  • [0275] 2900: Continuation of processes 2800
  • [0276] 2940: An updated sample space (18 flows) comprising sample space 2830 excluding labeled flows of the third sample (eight flows)
  • [0277] 2942: A fourth sample (fourth subset, 10 flows) drawn from the updated sample space 2940
  • [0278] 2944: A number (eight) of successfully labeled flows of the fourth sample
  • [0279] 2950: An updated sample space (ten flows) comprising sample space 2940 excluding labeled flows of the fourth sample (eight flows)
  • [0280] 2952: A fifth sample (fifth subset, ten flows) comprising the entire updated sample space 2950
  • [0281] 2954: A number (six) of successfully labeled flows of the fifth sample
  • [0282] 2960: An updated sample space (four flows) comprising sample space 2950 excluding labeled flows of the fifth sample (six flows)
  • [0283] 2962: A last sample (last subset, four flows) comprising the entire updated sample space 2960
  • [0284] 2964: A number (four) of successfully labeled flows of the last sample
  • [0285] 3000: Process of updating training data supplied to semi-supervised models 2745
  • [0286] 3010: An unlabeled flow
  • [0287] 3020: Initial sample space of current training epoch (40 flows)
  • [0288] 3030: Sample space (28 flows) after processing N₁ samples, N₁>0
  • [0289] 3032: A labeled flow
  • [0290] 3035: A set of labeled flows (12 flows) used to augment training data for semi-supervised models
  • [0291] 2745 after processing the N₁ samples (N₁ subsets)
  • [0292] 3040: Sample space (18 flows) after processing N₂ samples, N₂>N₁
  • [0293] 3045: A set of labeled flows (22 flows) used for training semi-supervised models 2745 after processing the N₂ samples (N₂ subsets)
  • [0294] 3100: Sequence of generating a suggested label for a flow
  • [0295] 3180: Potential new label suggestions in process 2760
  • [0296] 3190: Evaluation in processes 2760 and 2770
  • [0297] 3200: Processes of co-training employing two labelers (classifiers)
  • [0298] 3210: Training data containing both labeled and unlabeled traffic-flow records
  • [0299] 3212: A first labeler (classifier) based on packet-size parameters
  • [0300] 3214: A set of predicted flows of the first labeler
  • [0301] 3216: A process of updating training data 3210
  • [0302] 3219: A loop of successive activation of the first labeler to sequentially label unlabeled flow records
  • [0303] 3220: Training data containing both labeled and unlabeled traffic-flow records
  • [0304] 3222: A second labeler (classifier) based on packets arrival patterns (generally, time-domain packet-arrival patterns, including packets inter-arrival time patterns)
  • [0305] 3224: A set of predicted flows of the second labeler
  • [0306] 3226: A process of updating training data 3210
  • [0307] 3229: A loop of successive activation of the second labeler to sequentially label unlabeled flow records
  • [0308] 3260: A two-state switch
  • [0309] 3261: An input port for supplying output of labeler 3212 to switch 3260
  • [0310] 3262: An input port for supplying output of labeler 3222 to switch 3260
  • [0311] 3263: An output port for transferring switched flow records from labeler 3222 to update training data 3210
  • [0312] 3264: An output port for transferring switched flow records from labeler 3212 to update training data 3220
  • [0313] 3300: Composition of training data at start of labeling each sample, according to a first co-training method
  • [0314] 3310: Initial unlabeled records
  • [0315] 3316: Unlabelled flow records
  • [0316] 3320: Labeled flow records from the first labeler 3212
  • [0317] 3330: All flow records labeled at the first labeler 3212
  • [0318] 3335: Residual flow records after processing all samples in the first labeler 3212
  • [0319] 3350: Switched labeled flow records from the first labeler 3212
  • [0320] 3351: Switched labeled flow records from the second labeler 3222
  • [0321] 3356: Unlabelled flow records
  • [0322] 3370: Labeled flow records from the second labeler 3222
  • [0323] 3380: All flow records labeled at the first labeler 32212
  • [0324] 3385: Residual flow records after processing all samples in the second labeler 3222
  • [0325] 3400: Composition of training data at start of labeling each sample, according to a second co-training method
  • [0326] 3500: Successive, non-overlapping, training epochs
  • [0327] 3520: Cumulative training data over multiple training epochs
  • [0328] 3600: A first scheme of concurrent processes of training and real-time classification
  • [0329] 3640: Additional training data determined from classification of clients-provided flows
  • [0330] 3700: A second scheme for concurrent real-time classifications and updating training data
  • [0331] 3720: A process of acquiring predetermined training data to be used for synthesizing a real-time classification model for Epoch(0)
  • [0332] 3740: Concurrent operations during even-numbered epochs, starting with epoch(0)
  • [0333] 3741: Processes of receiving flow-classification requests from clients and determining a respective class for each flow
  • [0334] 3742: Processes of updating training data for an immediately succeeding epoch (an odd-numbered epoch) and synthesizing a real-time classification model to be used in processes 3761
  • [0335] 3750: Transition to subsequent-epoch (an odd-numbered epoch) processing
  • [0336] 3760: Concurrent operations during odd-numbered epochs, starting with epoch(1)
  • [0337] 3761: Processes of receiving flow-classification requests from clients and determining a respective class for each flow
  • [0338] 3762: Processes of updating training data for an immediately succeeding epoch (an even-numbered epoch) and synthesizing a real-time classification model to be used in processes 3741
  • [0339] 3770: Transition to subsequent-epoch (an even-numbered epoch) processing
  • [0340] 3780: An indication of continual activation of the loop of concurrent operations 3740 and concurrent operations 3760, necessitating continuous concurrent activation of the training component 110 and the real-time operational component 120 of Trafficwiz™ (FIG. 1).
  • [0341] 3800: Coupling processes of successive epochs to update training traffic records
  • [0342] 3810: A set of training traffic-flow records for a current epoch of index Π, Π≥0
  • [0343] 3820: Training-phase processes (labeling and modeling)
  • [0344] 3825: Traffic records received from users (clients) or the live network during Epoch(Q)
  • [0345] 3830: Real-time classification processes of Epoch(Q)
  • [0346] 3840: Processes of communicating classification results
  • [0347] 3850: A subset of the set of training flows, successfully labeled during Epoch(Q) in the training phase 3820
  • [0348] 3855: A process of age-filtering, applicable to Epochs succeeding the initial Epoch(0), where records of flows exceeding a prespecified age are excluded.
  • [0349] 3860: A process of retaining traffic-flow records and corresponding successful classifications to include in the set of training records of epoch(Q+1)
  • [0350] 3870: A process of combining age-filtered flow records from process 3855 and retained flow records from process 3860 to produce a set of training data for epoch(Q+1)
  • [0351] 3880: Set of training data 3880 for epoch(Q+1).
  • [0352] 3900: Exemplary outcome of the labeling stage for multiple training epochs
  • [0353] 3920: Labeled records used in the training phases of successive epochs, individually identified as 3920(0), 3920(1), etc.
  • [0354] 3940: Records of successfully classified flows corresponding to classification requests received from clients during successive epochs; individually identified as 3940(0), 3940(1), etc.
  • [0355] 3950: Composition of flow records after completion of an epoch, individually identified as 3950(0), 3950(1), etc., for Epoch(0), Epoch(1), etc., respectively.
  • [0356] 4000: Growth of training data
  • [0357] 4020: Cumulative counts of flow records labeled in the training phase
  • [0358] 4040: Cumulative counts 4040 of clients' flow records classified in the operational phase
  • [0359] 4100: Candidate Training data comprising raw unlabeled records, labeled records, and classified records
  • [0360] 4110: Unlabeled records
  • [0361] 4120: A mixture of labeled records and classified records
  • [0362] 4200: Modes of operation of the Trafficwiz™ appliance
  • [0363] 4210: Process of assembling training data at start of a current epoch
  • [0364] 4220: Steo of skipping the labeling stage if the training data does not contain unlabeled flow records
  • [0365] 4230: Process of activating the labeler
  • [0366] 4240: Process of synthesizing a classification modeler
  • [0367] 4250: Process of activating a real-time classification operative
  • [0368] 4300: Process of pruning training data
  • [0369] 4310: Training data at start of an Epoch
  • [0370] 4320: Cumulative training phase labeled records
  • [0371] 4330: A first pruning filter
  • [0372] 4340: Cumulative operational-phase classified records
  • [0373] 4350: A second pruning filter
  • [0374] 4400: An overview of a classifier
  • [0375] 4410: Unlabeled training flow data
  • [0376] 4420: A binary AI classifier module with integrated Receiver-Operating-Characteristics (ROC) sub-module
  • [0377] 4440: Non-payload flow data
  • [0378] 4450: Payload-carrying flow data
  • [0379] 4460: Multi-class AI model configured to predict flow traffic classes/categories and/or labels
  • [0380] 4500: Growth of a training dataset with unconstrained inclusion of training data acquired from the real-time classification stages
  • [0381] 4510: A point representing unconstrained cumulative training data size at a corresponding instant of time
  • [0382] 4520: Size (number of flow records) of initial training data
  • [0383] 4530: Size of training data extracted from the real-time classification stages of preceding epochs
  • [0384] 4600: Growth of a training dataset with constrained inclusion of training data acquired from the real-time classification stages according to a first policy of age filtering of training data
  • [0385] 4610: A point representing constrained cumulative training data size at a corresponding instant of time (under the first policy)
  • [0386] 4620: Declining size (number of flow records) of retained portion of the initial training data (of size 4520)
  • [0387] 4630: Size of age-filtered training data extracted from the real-time classification stages of preceding epochs
  • [0388] 4700: Growth of a training dataset with constrained inclusion of training data acquired from the real-time classification stages according to a second policy of age filtering of training data
  • [0389] 4710: A point representing constrained cumulative training data size at a corresponding instant of time (under the second policy)
  • [0390] 4730: Size of age-filtered training data extracted from the real-time classification stages of preceding epochs
  • [0391] 4800: Classification AI modeler components to generate an AI model for classification
  • [0392] 4805: A storage medium holding labeled AI training data
  • [0393] 4810: A data-cleansing module
  • [0394] 4820: A data imputation module
  • [0395] 4830: A module for removing flow records considered to be outliers
  • [0396] 4840: A data-transformation module
  • [0397] 4850: A feature-computation module
  • [0398] 4860: A feature-extraction module
  • [0399] 4870: A feature-selection module
  • [0400] 4880: A model-tuning module
  • [0401] 4890: A memory device holding an encoded AI model and specific parameters for use in real-time classification
  • [0402] 4900: Showing the processes and modules for Optimal selection of features used in AI model and real-time classifier
  • [0403] 5000: Closed-loop network, application and device traffic-flow classification
  • [0404] 5010: External AI training data
  • [0405] 5020: Clients' classification requests and/or live unclassified network traffic
  • [0406] 5030: Classified-flow data to clients
  • [0407] 5040: Classified-flow data to feedback to labeler for use as additional training data
  • [0408] 5100: A method of concurrent activation of labeler 760, synthesis of modeler 770, and real-time classification at classifier 780

[0409] The network appliance of the present invention extracts invariant features of encrypted traffic that are used to uniquely identify different applications. Such invariant features of encrypted traffic can be relied upon to pinpoint specific types of traffic without examining the contents of the traffic payload. AI models, whether Machine learning (ML) or Deep Learning (DL) models are trained on features extracted from representative traffic. The trained AI/ML models are then used for detection in future new traffic. It is understood that for the purposes of this document, the term AI is used to encompass ML, DL and related approaches, and may sometimes be used inter-changeably.

[0410] The premise behind the TrafficWiz ML-based Traffic Classification is that the network traffic generated by applications have unique temporal and spatial statistical properties, which are resilient to encryption and generally resilient to routine changes in application signaling protocols. If AI/ML systems are trained to learn the network behaviour for each class of applications, then the systems can be used to classify new network traffic based on the trained models. Further, this classification works accurately even if the traffic is encrypted. The statistical properties are reflected in statistical metrics extracted from the network traffic. Therefore, the classification also works even if there are minor changes in underlying protocol signaling behaviour for an application.

[0411] In AI/ML terminology, the statistical metrics which uniquely identify a particular application are referenced as “features”. These features are defined and extracted from network traffic flows. There are literally thousands of features which can be defined and tracked for a given traffic flow. However, not all features can facilitate robust and accurate AI classification models nor can all features facilitate real-time operation of the AI model classifiers on requisite high speed networks. Examples of such features are metrics associated with packet inter-arrival times distribution, byte counts, packet counts, flow duration, as well as more sophisticated metrics such as entropy (measure of randomness).

[0412] Once an artificial-intelligence (AI) model has been developed, TrafficWiz™ utilizes the model to examine and classify incoming encrypted network traffic and identify the underlying applications which generated the traffic.

[0413] A number of steps are required to develop an ML model for the TrafficWiz toolset, which would robustly encapsulate and represent the behaviour of network traffic, and suitably classify network traffic with a specified accuracy.

[0414] FIG. 1 is a schematic 100 of a network appliance, herein referenced as Trafficwiz™, for traffic-flow characterization implemented as an integrated appliance. A traffic flow is hereinafter referenced as a “flow” for brevity.

[0415] The integrated appliance comprises a training component 110 activated during a training phase, an operational component 120 activated in real-time in a continuous operational phase, and a module 130 for updating the training data based on outcome of the operational phase. The training component produces a specific ML model and associated parameters 116 to be used for real-time classification in component 120.

[0416] A memory device 112 stores initial global training data which may be totally unlabeled or partially labeled. The training data is organized into records of individual flows. As illustrated in FIG. 10 to FIG. 13, to be described below, a record of a flow indicates a unique identifier and a respective superset of features (also referenced as “descriptors”, “metadata” or “statistics”. A superset of features may be reduced to a representative set of prominent features. A memory device 114 holds the initial global training data which is gradually updated to include training data that may be extracted during the operational phase.

[0417] The operational component 120 is configured to receive flow classification requests 121 from users (also referenced as “clients”, FIG. 6) of appliance 100, or from live network traffic and produce respective per-flow classifications. The operational phase forms responses 122, to classification requests 121 whether from users or from live network traffic, and communicates the responses through a network interface illustrated in FIG. 3. The produced individual, per-flow, classifications are also supplied to module 130; data 123 relevant to classified flows defined in requests 121 is supplied to module 130 for continually updating the training data and restructuring the training data stored in memory device 114.

[0418] In a preferred implementation, a continuous recursive training-data procedure 150 is performed during successive predefined training epochs, activating the training component 110, the operational component 120, and module 130 leading to a continually improved updating training data of memory device 114 of gradually increased proportion of labeled flows (labeled records).

[0419] FIG. 2 illustrates a partitioned arrangement 200 of the network appliance of FIG. 1 functioning as a partitioned appliance (a partitioned Trafficwiz™). The partitioned appliance comprises a stand-alone training module 230 (a training part of integrated network appliance 100) and a stand-alone operational module 240 (a classification part of integrated network appliance 100). The stand-alone training module 230 and the stand-alone operational module 240 are hardware entities, as illustrated in FIG. 4) which may be collocated or geographically separated. Modules 230 and 240 may be interconnected directly, or through a dual path in data network 220. The dual path may be a dedicated path or a switched path through a network.

[0420] FIG. 3 illustrates main hardware components 300 of the appliance of FIG. 1 implemented as an integrated appliance. The appliance communicates with sources of information and users of the appliance through network interfaces. A training module of network appliance 100 communicates through a respective network interface 320 and a dual channel 321 to a data network. An operational module of network appliance 100 communicates through a respective network interface 325 and a dual channel 326 to the data network. A memory device 330 stores training instructions specific to the training functions. A memory device 335 stores software instructions specific to the real-time classification functions. A memory device 340 stores current training data (initially storing pre-determined training data of memory device 112). A memory device 342 stores records of training flows (training records) that were not labelled during a training epoch for use in further analyses. A memory device 345 stores additional flow records acquired during the operational phase (real-time classification phase) of a training epoch to be considered as candidate training data for use in a subsequent training epoch. A processor 350 is coupled to the network interface 320, network interface 325, memory devices 330, 335, 340, 342, and 345. Processor 350 is generally implemented as an assembly of hardware processing units operated in parallel and/or in a pipelined arrangement.

[0421] FIG. 4 illustrates hardware implementation 400 of the appliance of FIG. 2 in the form of two stand-alone hardware entities that may be coupled directly or through a data network. The stand-alone training module 230 is implemented as a hardware structure 430. The stand-alone operational module is implemented as a hardware structure 440.

[0422] FIG. 5 illustrates a spatially distributed appliance 500 for flows characterization comprising a central training module 530, similar to training module 230 (implementation 430), and multiple operational modules 550, individually 550(1), 550(2), etc., each operational module 550 being similar to module 240 (implementation 440) of the partitioned arrangement 200 (implementation 440). The central training module 530 comprises a network interface 532, similar to network interface 320 of FIG. 3. The multiple operational modules 240 (implementation 440) that may be geographically distributed and connected in a star configuration to the central training module 530 through a data network. The central training module 230 distributes training results 510 to the operational modules and receives results 520 (individually 520(1), 520(2), etc.) of real-time classifications from the operational modules 550. Each operational module 550 comprises a respective operational-module network interface 552.

[0423] FIG. 6 illustrates a system 600 of integrated and distributed Trafficwiz™ appliances coupled to data network 220. Two integrated Trafficwiz™ appliances 100, as well as a distributed appliance 500 are communicatively coupled to a plurality of clients 610 through the data network. The distributed appliance comprises a stand-alone training module 530 communicatively coupled to multiple stand-alone operational modules 550 through the network. Each operational stand-alone module 550 is configured to communicate with the plurality of clients 610 through the data network.

[0424] FIG. 7 is an overview 700 of the main processes 720 performed by a Trafficwiz™ appliance, whether integrated, partitioned, or distributed, comprising training-phase processes 740 and operational-phase real-time processes 750. Processes of the training phase 740 comprise labeling of unlabeled flow records performed at a labeler 760 and synthesizing an AI model by a Modeler 770 to be used by the classifier 780 in the operational phase. Processes of the operational phase 750 comprise classifying fresh traffic flows received from the live network or from clients using the classifier 780.

[0425] FIG. 8 illustrates more detailed basic processes 800 performed at a Trafficwiz™ appliance whether integrated (FIG. 1), partitioned (FIG. 2), or distributed (FIG. 5).

Process 820 acquires set of traffic records for training (comprising an initial predetermined set of traffic records and any traffic records labeled during the real-time operational phase). Process 830 acquires or specifies a set of recognizable traffic-flow classifications. Process 822 attempts to label each of the traffic records. Process 840 synthesizes a traffic-flow modeler based on successfully labeled traffic-flow records. Process 850 synthesizes a real-time traffic-flow classifier. Process 860 continually receives fresh flow data from clients or live network traffic 610 through network interface 552 (FIG. 5). Process 880 performs real-time traffic-flow classifications using synthesized classifier 850 to produce respective flow classifications 890.

[0426] FIG. 9 illustrates an exemplary predefined set 900 of flows comprising homogeneous flows, as well as a predefined set of composite flows each formed as a multiplex of a respective subset of the set of homogeneous traffic flows.

[0427] The exemplary set of homogeneous flows (indexed as 01 to 10, comprises: an audio stream (a one-directional, downstream flow); an audio conversation (colloquially “audio chat”); a one-directional text; a two directional text; a file transfer; an e-mail message; a point-to-point communication; a web browsing session; a video-stream of a first type (e.g., low resolution, a specific family of protocols; etc.); and a video stream of a second type (e.g., high-definition video streams).

[0428] Given the continuously evolving communication services, a set of composite flows, each encoded as a multiplex of homogeneous flows may be considered. For example, a typical video stream comprises, in addition to video signals, audio signals and -encoded texts.

[0429] FIG. 10 illustrates an exemplary training dataset 1000 of flows to be labeled, according to the predefined flow classes, undergoing preparatory (training) operations of filtering, clustering, and sampling. An initially acquired set 1010 of unlabelled flows, stored in a respective memory device, comprises records 1012 of individual flows which may include outliers 1014 which would be filtered out.

[0430] A process 1020 filters the flow records of the set 1010 of unlabeled flows according to pre-established criteria which may be periodically or occasionally updated as the need arises. The filtering process produces a set 1022 of flow records retained to be used in the processes of the training phase. Process 1030 clusters and visualizes the flow records 1022 using any of well-established techniques. Four clusters of records, referenced as 1032(1) to 1032(4) are illustrated.

[0431] Due to feasibility considerations, labeling of set 1022 of filtered flows is performed using successive subsets of set 1022, selected according to an appropriate sampling method. The selected subsets are also referenced as “samples”. Set 1022 is a sampling space which is modified after drawing and processing each sample.

[0432] Process 1040 samples set 1022 of filtered flows to extract subsets of a number of flow records within a specified range of the number of flows per sample. In the illustrated example, six samples (six subsets) of set 1022, individually labeled as samples 1042(1) to 1042(6). Each sample may include flows belonging to different clusters. As indicated in the exemplary case of FIG. 10, each sample includes flows from each of the four produced four clusters. Conversely, each of the four clusters is represented in each of the six samples.

[0433] It is important to note that the samples are drawn during successive time periods and the subset of flows in each sample may not be entirely labeled. Labeled flows of each sample are removed from set 1022 of filtered flows while unlabeled flows of a sample remain in the sampling space. Thus, successive samples may include common flows. The intersection of successive samples is not illustrated in FIG. 10 but will be elaborated in other drawings; see, for example, FIGS. 28 to 30.

[0434] Generally, a large filtered set of flow records may be processed for labeling in numerous samples (numerous subsets or flow records). The number of successive subsets to be drawn during a training epoch may be predefined or determined according to a criterion such as exhaustion of the training dataset of a current training epoch or labeling saturation. Labeling saturation occurs if a predefined number of successive samples fail to increase the number of labeled flow records, i.e., the remaining sample space does not change. This condition may arise when the set of filtered flow records 1022, which excludes outliers, may contain flows that cannot decisively be associated with any of the pre-defined flow classes.

[0435] FIG. 11 illustrates an example 1100 of flow characterization. An identifier 1121 and specific characteristics 1122 are associate with each flow 1120. The identifier 1121 is preferably expressed as a tuple 1130 containing at least five entries {Source address; destination address; source port; destination port; protocol according to which the flow is structured}. The characteristics 1122 of a flow are preferably organized as a table 1140 of a superset of predefined generic features, each feature associated with a respective flow-direction code 1150. The features are also referenced as “descriptors”, “metadata” or “statistics”.

[0436] The values 1151 of code 1150 indicate flow-direction pertinence of respective features. For example, a code of “00” indicates that a corresponding feature is not relevant to a specific flow 1120 under consideration, a code of “01” indicates that a respective feature is applicable in the upstream (forward) direction only, a code of “10” indicates that a respective feature is applicable in the downstream (backward) direction only, and a code of “11”, indicates a respective feature is applicable in both the upstream and downstream directions. Feature specifics 1160 include at least a respective index 1161 and a respective coded definition of each feature.

[0437] The flow features are captured for each direction, i.e., forward and backward, and comprise:

• • [0438] (1) Four features capturing Minimum, mean, maximum, standard deviation of packet sizes;
  • [0439] (2) Four features capturing Minimum, mean, maximum, standard deviation of packet inter-arrival time;
  • [0440] (3) Two features capturing total number of packets;
  • [0441] (4) Two features capturing total bytes in a flow;
  • [0442] (5) Two features capturing total payloads in a flow;
  • [0443] (6) Two features capturing total number of payload packets in a flow;
  • [0444] (7) Four features capturing entropy of packet sizes and entropy of inter-packet arrival time;
  • [0445] (8) Fifty features capturing probability of state transition between consecutive packets with same and different packet sizes;
  • [0446] (9) Fifty features capturing probability of state transition between consecutive packets with same and different inter-packet arrival time; and
  • [0447] (10) Four feature capturing probability of state transition for packet direction between consecutive packets.
  • [0448] These features are key to representing temporal and spatial characteristics of encrypted traffic. It is understood that additional features can be defined and added to the above set of features.

[0449] The superset of features may be pruned to a smaller set of features without sacrificing labeling accuracy as illustrated in FIG. 13.

[0450] FIG. 12 illustrates a process 1200 of sorting the predefined superset 1140 of flow features according to a predefined set of feature categories to facilitate a process of co-training to be discussed below. To start, a predefined set 1210 of Γ feature categories, Γ>1, individually referenced as 1212(1) to 1212(Γ), is acquired. A process 1240 sorts features of the superset 1140 of features according to the predefined set 1210 of feature categories to produce Γ bins 1260 of category-specific features, the bins are individually referenced as 1260(1) to 1260(Γ). Each bin includes indices 1161 of features of the superset of features 1140. Separately, the grouping of features can be utilized as the basis of generating different AI models for classification.

[0451] FIG. 13 illustrates a process 1300 of optimal selection of a set of prominent features from the predefined superset of features 1140. Process 1320 illustrates one embodiment of optimal selection which adapts feature data for processing in a genetic-algorithm module 1330. The genetic algorithm is activated to produce a set 1340 of prominent features extracted from the predefined superset of features.

[0452] FIG. 14 illustrates a scheme 1400 for continuous training of the Trafficwiz™ appliance (whether integrated 100, partitioned 200, or distributed 500). A process 1410 continually acquires global training flow records (global training dataset) over a moving time window, such as a moving training epoch of a predetermined duration (a week, a month, etc.). In a training stage 1420 for a current time-window (a current epoch), processes 1430 selects successive samples (subsets) of the global training flow records of the current time-window. Each of processes 1440(0) to 1440(S*), S*being an index of a last sample of the current time window, attempts to label each constituent flow. The labeling processing may conclude with successfully labeling all, or some, flows of the global training dataset. The successfully labeled flows are used in process 1460 to synthesize the modeler 770, i.e., generate the model and associate algorithmic determine parameters defining the operational modeler 770.

[0453] FIG. 15 illustrates nesting arrangement 1500 of stages for the scheme of FIG. 14, where each training phase 1510 of a continual sequence of training epochs includes multiple preprocessing stages 1520, each preprocessing stage including multiple sampling stages 1530, each sampling stage including multiple single-flow processing stages 1540 (referenced as Kernel stages).

[0454] FIG. 16 further details the stages of FIG. 15 indicating recursive processes 1600 of updating of the global training dataset. Processes of outer loop 1610 are activated during successive time windows (successive training epochs) to produce refreshed training data 1612 for use in a current time window (a current training epoch). Processes 1620 of intermediate loop 1620 selects successive subsets 1614 of the training dataset executed during successive sampling periods, i.e., periods of attempting to label successive subsets of traffic flows. Processing 1630 of the inner loop 1630 selects a flow of one traffic flow, with corresponding data 1616, and attempts to label the flow.

[0455] FIG. 17 illustrates a method 1700 of labeling a massive set of unlabeled or tentatively labeled traffic-flow records, based on progressively modifying the composition of the massive set as traffic-flow records get labeled.

[0456] Initially, a memory device 1710 stores a superset of M unlabeled flow records, M may be a relatively large integer. To start, a process 1720 extracts a sample of at most a predefined number, L, L<<M, of flow records from the unlabeled flow records stored in memory device 1710. The sample is presented to process 1750 which carries out a labeling process utilizing multiple approaches including AI and machine-learning approaches, limited to handle at most L records at a time. After attempting to label all flows of the sample, process 1760 places records of successfully labeled flows in a memory device 1780 and updates the content of memory device 1710 to remove records of the successfully labeled flows. Process 1730 then determines whether another sample need be drawn.

[0457] Recursive process 1740 cyclically activates processes 1720, 1750, 1760, and 1730 until a termination condition is reached.

[0458] Process 1730 revisits process 1720 to extract another sample if none of predetermined conditions is reached. The predetermined conditions may include:

• • [0459] (1) the number of remaining unlabeled flow records in memory 1710 is less than a respective predetermined lower bound;
  • [0460] (2) the proportion of the initial total number, M, of unlabeled records that has been successfully labeled exceeds a respective predetermined upper bound;
  • [0461] (3) a succession of a predetermined number of samples failed to produce any labeled flow records (which may trigger an alarm signal); and
  • [0462] (4) any criterion for ending the recursive process 1740 has been attained.
  • [0463] If any of the four conditions is reached, process 1790 terminates the labeling process for a current time window (a current epoch).

[0464] FIG. 18 is a flow chart 1800 of a specific implementation of the method of FIG. 17 In process 1810, a file, referenced as a “label file”, for holding labeled records of traffic flows, is initialized as an empty file. Process 1820 assembles records of candidate training data in a “global file”. For a first training epoch (training time window), the global file contains only unlabeled records. For subsequent training epochs, the global file may contain a mixture of prior records as well as newly acquired records. At branching process 1830, loop 1835 is activated if a condition for continuing processing of a current epoch is met; otherwise training data is prepared for a new epoch (new time window) in process 1890 then process 1820 is revisited.

[0465] Loop 1835 activates processes 1840, 1850, 1860, 1865, 1870, and 1880 before returning to branching process 1830:

• • [0466] process 1840 selects a sample (a subset) of a preselected number of records from the global file;
  • [0467] Branching process 1845 proceeds to process 1850 if the label file contains a sufficient number of labeled flow records, otherwise 1845 proceeds directly to process 1860 where labeling is attempted without the benefits of the semi-supervised learning model;
  • [0468] process 1850 trains a semi-supervised learning model based the label file;
  • [0469] process 1860 attempts to label each record of the selected sample;
  • [0470] process 1865 attempts to label each unlabeled record of the sample using available suggestions;
  • [0471] process 1870 inserts successfully labeled records into the label file;
  • [0472] process 1880 removes successfully labeled records of the sample from the global file then returns to branching process 1830.

[0473] FIG. 19 illustrates an exemplary outcome 1900 of applying the implementation of FIG. 18, indicating growth of the number of labeled records during successively selected subsets of the training dataset. Initially, the global file contains a number M, reference 1901, of unlabeled flow records and the label file contains zero labeled flow records (reference 1902). After processing a number a (at least one) of seed samples to build-up the label file to a prerequisite level, subsequent samples are drawn from the global file to be processed with the benefit of the semi-supervised learning module (process 1860). Counts of unlabeled records after processing successive samples, starting with the last seed sample, are individually referenced as 1920(α+j), 0≤j≤j*, where j* is determined at the branching process 1830. Counts of labeled records after processing successive samples, starting with the last seed sample, are individually referenced as 1940(α+j), 0≤j≤j. Counts of all records after processing successive samples, starting with the last seed sample, are individually referenced as 1950(α+j), 0≤j≤j. A count (reference 1960) of residual flow records of the global file is indicated. The residual flow records are flow records that were not labeled, due to reaching any of conditions (1) to (4) recited above in the description of FIG. 17.

[0474] Thus, residual flow records contain unlabeled flows, which may belong to legitimate network traffic, or may contain malicious code which does not fit into traffic classification types. Therefore filtering residual flows may improve computer security, by separating the malicious traffic from labeled classified traffic flows.

[0475] FIG. 20 illustrates labeling attainment 2000 for successive samples based on growth of the overall number of trained flow records as successive samples are drawn from the global file and processed for labeling. The sample sizes 2020 are individually referenced as 2020(Θ+j), 0≤j≤j*, where j* is determined at the branching process 1830. The number 2020(k) of drawn flow records of a sample of index k, Θ≤k≤(Θ+j*), is denoted φ_k. The numbers of labeled and unlabeled flow records are denoted η_k and λ_k, respectively: η_k+λ_k=φ_k.

[0476] As illustrated in FIG. 18, when the cumulative number of labeled flow records is smaller than a predefined threshold, labeling of pending unlabeled flows is performed without the benefit of the semi-supervised learning model. As more flows are successfully labeled, the probability of successful labeling in a single attempt increases leading to an increase in the ratio (η_k/φ_k) over successive samples. As illustrated, the ratio (η₄/φ₄) is significantly larger than the ratio (η₁/φ₁).

[0477] FIG. 21 illustrates an exemplary global set 2100 of flows, where each flow is presented as a vector 2120 of a multi-dimensional space. A simplified two-dimensional space is used for ease of presentation. A vector 2120 of flow characteristics is also referenced as a flow record, or simply a “record”, and illustrated as a dot in FIG. 21. The global set of flows may be divided into a set of clusters, each cluster containing a number of records. Several clustering methods may be employed.

[0478] In a popular clustering method known as the k-mean method, a number of clusters is specified, each initialized to have a proposed centroid, which may be selected from the global set 2100, then clusters are formed according to proximity-to-centroid criteria. With the number of centroids specified, the resulting cluster sizes (numbers of flows within each cluster) may vary significantly and iterative revisions may be applied to redistribute flows of clusters of sizes outside an acceptable range.

[0479] In another popular clustering method known as the DBSCAN method (density-based spatial clustering of applications with noise), two parameters, ρ and ν, are specified where ρ denotes a radius of a hypersphere and ν denotes a minimum number of vectors within a hypersphere of radius ρ. Both ρ and ν are scalars. For a given global set of records (vectors), specified values of a pair {ρ, ν} leads to a respective number of clusters. The labeling process (FIG. 7, FIG. 10. FIG. 15) relies on clustering the global set of vectors with a preferred number of clusters in the range 20 to 25. Thus, if the DBSCAN method is employed, the parameters {ρ, ν} have to be selected to produce the desired number of clusters. There may be numerous paired values that produce a target number of clusters. Starting with a lower bound ρ₁ of ρ and an upper bound ρ₂ of ρ, corresponding values of ν₁ and ν₂, respectively of ν may be determined based on a predefined function of ρ. A value of ρ that yields the target number of clusters may be determine using well known methods such as successive bisection, successive linear interpolation, or a combination of linear interpolation and parabolic interpolation. Two hyperspheres 2140(1) and 2140(2), of radii ρ₁ and ρ₂, are illustrated. Each of the corresponding encompassed numbers of vectors, ν₁ and ν₂, exceeds or equals a corresponding lower bound.

[0480] FIG. 22 illustrates an example 2200 of clusters C1, C2, C3, C4, and C5 of the global set of flows. With appropriate random selection of flows of each sample (each subset of flows), flows of each cluster would be selected in several samples.

[0481] FIG. 23 illustrates an example 2300 of successive samples S₁, S₂, . . . , S₈, (reference 2320(1) to 2320(8)) drawn from the global set 2100 of flows, with partial replacement where flows of any sample that could not be decisively labeled are returned to the global set. As illustrated in FIG. 20, the proportion of successfully labeled flows within each sample grows steadily in successive samples because any flows not labeled during the pass of a particular sample will be considered for labeling in the pass of a future sample. The number λ_k of unlabeled flows in a sample of index k, k≥0, remains in the global file. In the illustrated example, the first drawn sample S₁ contains 25 unlabeled flows. selected from the initial flows present in the global file, of which several flows were not successfully labeled and remained in the global file. The second sample S₂ drawn from the reduced global set 2100 contains 30 unlabeled flows of which 5 flows were considered in sample S₁. The formation of successive samples continues with each sample potentially containing a mixture of flows which were not considered yet as well as flows that may have been considered, but are still unlabeled, in the processing of preceding samples.

[0482] Appropriate random selection of constituents of each sample is of paramount importance.

[0483] FIG. 24 illustrates a process 2400 of sampling with partial replacement. A miniature global file containing only 40 flows is clustered into five clusters 2410(1) to 2410(5), denoted C₁ to C₅, respectively. Following the process of FIG. 23, eight samples 2420(1) to 2420(8), denoted S₁ to S₅, respectively are formed. The combined content of the eight samples is larger then 40 because several flows were not labeled in a first attempt of a respective sample-processing stage. The sample of index 18 of the clustered flows was selected once and successfully labeled in sample S₂. However, the sample of index 27 was selected in samples S₄, S₅, and S₈, where it was successfully labeled in the latter.

[0484] FIG. 25 illustrates processes 2500 of the pre-processing stage. Process 2510 acquires raw unlabeled flow records to be stored in a respective memory device 2512. Process 2520 filters the raw unlabeled flow records to exclude flow records considered to be outliers according to respective methods. Resulting filtered raw unlabeled flow records may be stored in a separate memory device 2522 or in a partition of memory device 2512. The filtered raw unlabeled data is supplied to a clustering process 2530 and to a process 2580 for generating augmented flow records.

[0485] Process 2530 performs a clustering process of the filtered raw unlabeled flow records, stored in memory device 2522, according to any of known techniques as discussed above with reference to FIG. 23. Relevant clustering data, indicating flow records of each cluster, is stored in a memory device 2532. Process 2534 generates a flow-scatter plot and a packet-sequence plot.

[0486] Process 2550 produces a DPI (Deep packet inspection) suggested label for each unlabelled flow. It is noted that DPI of encrypted packets relies on metadata.

[0487] Process 2555 produces a suggested label for each unlabelled flow based on a pre-trained ML/DL (machine learning/deep learning) model. Process 2560 produces a suggested label for each unlabelled flow based on specific rules. Process 2565 produces a suggested label based on traffic patterns

[0488] Process 2580 processes the filtered raw unlabeled data stored in memory device 2522 to produce augmented filtered raw unlabeled flow data indicating, for each unlabeled flow, the suggested labels produced in processes 2550, 2555, 2560, and 2565.

[0489] The augmented filtered raw unlabeled flow data is stored in a memory device 2585 to be used in the labeling process.

[0490] FIG. 26 illustrates processes 2600 of the sampling stage 1530, 1614. Process 2610 initializes a set of labeled flows as an empty set. Branching process 2620 determines completion, or otherwise, of processing unlabeled flows of a current training epoch. If completion condition of a current epoch is met, process 2620 branches to process 2680 to deliver results to processes 770 (FIG. 7) of deriving operational models and start a new epoch. Otherwise, process 2620 branches to process 2640 to select a sample (a subset) of the global set of pending augmented unlabeled records and submit individual flow records of the sample to the kernel module. The Kernel module returns, for each unlabeled flow, either a respective label or an indication of unsuccessful labeling.

[0491] Recursion 2650 selects samples of the augmented unlabeled flows for processing in the kernel module until a completion condition for a current training epoch is met.

Process 2670 updates a set of cumulative labeled flows resulting from processing individual flows of a current sample. As in the generic method of FIG. 17, The completion conditions may include:

• • [0492] (a) the number of remaining unlabeled flow records in the global set 2585 of pending augmented unlabeled flows is less than a respective predetermined lower bound;
  • [0493] (b) the proportion of the initial total number of unlabeled records that has been successfully labeled exceeds or equals a respective predetermined target;
  • [0494] (3) a succession of a predetermined number of samples fails to produce any labeled flow records (which may trigger an alarm signal); and
  • [0495] (4) any other criterion for ending the recursive process 2650 has been attained.

[0496] FIG. 27 illustrates processes 2700 of the kernel stage 1540, 1616. Branching process 2720 receives flow records of a current sample from process 2640 of sampling stage 2600 as well as results of processing of individual records of the current sample and determines continuity or completion of processing records of the current sample. If completion is determined, labeling results 2795 are communicated to the sampling stage 2600. Otherwise, the loop 2730 of processes 2725, 2740, 2745, 2750, 2760, 2765, and back to 2720 is followed until process 2720 determines completion of considering each record of the sample.

[0497] Process 2725 selects an unlabeled flow of the current sample. Branching process 2740 determines whether the cumulative number of labeled records of the current epoch is sufficient to train a semi-supervised model. If the cumulative number of labeled records is below a respective specified lower bound, process 2750 is activated, bypassing process 2745; otherwise process 2745 is activated to train semi-supervised models and suggest a label accordingly. One example of the semi-supervised model is based on co-training algorithm that will be described in more detail with regard to FIG. 32. AI ternatively, a label spreading algorithm may be used for semi-supervised models for suggesting labels in the process 2745, as cited in the information disclosure statement. In the description of FIG. 27 it is understood that in some embodiments multiple flows may be selected in 2725 in order to enhance performance, in which case all corresponding steps in FIG. 27 would thus process a set of flows in batch instead of one at a time.

[0498] Process 2750 produces a system-suggested label based on all produced label suggestions of the current flow (with or without the suggested label of process 2745). Process 2760 verifies, and may modify, the system suggested label. Process 2760 employs an automaton to review the system-suggested label; the automaton may be a human analyst.

[0499] Branching process 2765 returns to process 2720 without further processing if process 2760 rejects, and does not modify, the system suggested for a current flow (thus, the current flow remains unlabeled) or proceeds to a further process 2770 of machine evaluation.

[0500] Branching process 2775 returns to process 2720 without further processing if process 2770 rejects the outcome of process 2760, despite approval or modification in process 2765, thus the current flow remains unlabeled. Otherwise, the current flow is labeled and transferred from the global file to the label file in process 2780. Process 2780 further removes the current labeled flows from the global file.

[0501] FIG. 28 illustrates intermediate results 2800 of the labeling stages of FIG. 26 and FIG. 27 for an exemplary case of a miniature global set of 40 flows representing the initial sample space 2810. FIG. 29 is a continuation 2900 of FIG. 28.

[0502] A first sample 2812 of ten flows of sample space 2810 is drawn of which six flows are successfully labeled (reference 2814). With six flows being labeled, the initial sample space 2810 is updated to sample space 2820 of 34 flows. A second sample 2822 of ten flows is drawn from the updated sample space 2820 of which eight flows (reference 2824) are successfully labeled. With a total of 14 flows being labeled, the sample space is updated to sample space 2830 which includes 26 unlabeled flows. A third sample 2832 of ten flows is drawn from the updated sample space 2830 of which eight flows (reference 2834) are successfully labeled. With a total of 22 flows being labeled, the remaining unlabeled flows (reference 2835) constitute a current updated sample space of 18 flows.

[0503] A fourth sample 2942 of 10 flows is drawn from the updated sample space 2940 of which eight flows (reference 2944) are successfully labeled, yielding an updated sample space 2950 of ten flows. A fifth sample 2952 comprises the entire updated sample space 2950 of which six samples (reference 2954) are successfully labeled yielding an updated sample space 2960 of four flows. A sixth sample 2962 comprises the entire updated sample space 2960 all of which are successfully labeled (reference 2964).

[0504] FIG. 30 illustrates a process 3000 of updating training data supplied to semi-supervised models 2745. Initially, for a current epoch, the global file contains a set 3020 of unlabeled flows 3010. The set 3020 is a sample space of the current epoch (40 flows). After processing a number, N₁, of samples, N₁>0, the label file contains a set 3032 of labeled flows (12 flows) and the global file contains a reduced number (28) of unlabeled flows constituting updated sample space 3030. If the label file contains a number of labeled flows that at least equals a predefined threshold, the set 3032 would be used for training semi-supervised models 2745. After processing N₂ samples, N₂>N₁, the label file contains a larger set 3045 of labeled flows (22 flows) and the updated sample space 3040 contains a reduced number (18) of unlabeled flows. Set 3045 is then used for training semi-supervised models 2745.

[0505] FIG. 31 illustrates a sequence 3100 of generating a suggested label for a flow. Modules 2550, 2555, 2560, 2565 are described above with reference to FIG. 25. Modules 2745 and 2750 are defined in the description of FIG. 27. Process 2760 may suggest a new label (reference 3180). Evaluations (reference 3190) are performed in processes 2760 and 2770

[0506] In one embodiment, a label may be selected among multiple suggested labels using a voting method or a weighted voting method, where different weights are assigned to different suggested labels, please refer for example to an article in Wikipedia on “Weighted voting” cited in the Information Disclosure Statement.

[0507] AI ternatively, a Bayesian approach may be used for selecting a label among multiple suggested labels, to summarize the prediction result of each suggestor and give the agent the most reliable suggestion.

[0508] In one embodiment, the weight associated with the quality of pre-trained suggestors is determined either offline apriori using a system with pre-determined ground truth or utilizing the inherent expertise of the automaton. In either case, the system will utilize the most reliable suggestor's advice based on weights assigned to their suggestion. Each time the automaton/analyst accepts a suggested label, the Beta distribution of each suggestor is updated

[0509] In one example, all the suggestors carry out a prediction for 50 flows, and are reviewed by the automaton/analyst, where α represents the number of correct predictions, and β represents the number of incorrect predictions:

• • [0510] Suggestor A: α1=40, β1=10
  • [0511] Suggestor B: α2=30, β2=20
  • [0512] Suggestor C: α3=10, β3=40

[0513] As a result of the above, each suggestor will have a Beta distribution of beta (α, β), which is used to calculate the Posterior Distribution. The system will select the label predicted by the suggestor whose individual posterior has the highest mean as the most reliable. It will rely on this label as the system suggested label.

[0514] After the first round of labeling, the analyst feedback will result in a continuously updated beta distribution for each suggestor. If a suggestor predicted label is the same as the automator/analyst predicted label, the number of correct predictions will add up to α, the number of incorrect predictions will add up to β:

• • [0515] Suggestor A: α1=40+5, β1=10+5
  • [0516] Suggestor B: α2=30+9, β2=20+1
  • [0517] Suggestor C: α3=10+1, β3=40+9

[0518] It is understood that although the above example is presented with a single semi-supervised learning suggestor, that one embodiment of the system could include two or more semi-supervised suggestors.

[0519] FIG. 32 illustrates processes 3200 of co-training employing two labelers (classifiers). A first labeler 3212 utilizes an AI model with features representing packet-size parameters. A second labeler 3222 utilizes an AI model with features representing packet inter-arrival times/patterns. The first labeler uses training dataset 3210 which is continually updated as more flow records are labeled. Likewise, the second labeler uses a training dataset 3220 which is continually updated as more flow records are labeled. Initially, training datasets 3210 and 3220 are identical.

[0520] FIG. 33 illustrates unlabeled flow records 3316, gradually labeled flow records 3320 of training data set 3210, unlabeled flow records 3356, and gradually labeled flow records 3370 of training data set 3220.

[0521] Gradually labeled flow records 3320 of dataset 3210 are held in a storage medium 3214. Gradually labeled flow records 3370 of dataset 3220 are held in a storage medium 3224. Process 3216 updates training dataset 3210 as more unlabeled flow records change status from unlabeled to labeled. Likewise, process 3226 updates training dataset 3220 as more unlabeled flow records change status from unlabeled to labeled. Processing loop 3219 repeatedly activates labeler 3212 and updates dataset 3210 until a respective completion condition is met. Likewise, processing loop 3229 repeatedly activates labeler 3222 and updates dataset 3220 until a respective completion condition is met. The completion condition maybe any of: (1) all, or a prescribed proportion, of flow records have been labeled, (2) a labeler fails to increase the count of labeled records for a specified number of activations, indicating that remaining unlabeled flow records cannot be labeled.

[0522] A switch 3260 has two input ports 3261 and 3262 for supplying outputs of label predictors/suggestors 3212 and 3222, respectively to the switch. The switch has two output ports 3263 and 3264 for transferring switched flow records from labeler 3222 to update training data 3210 and transferring switched flow records from labeler 3212 to update training data 3220, respectively.

[0523] Switch 3260 connects input port 3261 to output port 3264 to provide labeled flow records produced at labeler 3212 to updating process 3226. The switch connects input port 3262 to output port 3263 to provide labeled flow records produced at labeler 3222 to updating process 3216. Thus, contents of labeled flow records held in storage medium 3214 are used to update both data sets 3210 and 3220. Likewise, contents of labeled flow records held in storage medium 3224 are used to update both data sets 3220 and 3210. The result of switching the unlabeled records between the two training models is to increase the likelihood that a flow label is predicted. In other words, if one label suggestion model cannot predict/suggest a label with high confidence, it is possible that by switching the unlabeled records to be predicted between the two models, that a desired result can be achieved. The exchange labeled flow records may be done according to different policies as illustrated in FIG. 33 and FIG. 34.

[0524] FIG. 33 illustrates composition 3300 of training data at start of labeling each sample of sampling stage according to a first co-training policy for a case of co-training where both of the candidate training data sets from this sample 3210 and 3220 initially contains unlabeled records 3310.

[0525] As illustrated, the output of the first labeler (labeler-A, also referenced as classifier-A) at the start of processing each sample, beyond the first sample which is labeled without the help of labeled flow records, comprises labeled records 3320(λ). λ≥0. Labeled records 3320(0) are produced during processing of sample(0) and used to label sample(1). Labeled records 3320(1) are produced during processing of sample(1) and used to label sample(2), and so on. Storage medium 3214 holds cumulative labeled records 3320(·). The count of cumulative labeled flow records 3330 after processing the last sample in the first labeler may be less than the count of the initial unlabeled flow records, the difference, referenced as residue 3335, may be further inspected for reasons other than labeling.

[0526] The output of the second labeler (labeler-B, also referenced as classifier-B) at the start of processing each sample, beyond the first sample which is labeled without the help of labeled flow records, comprises labeled records 3370(λ). λ≥0. Labeled records 3370(0) are produced during processing of sample(0) and used to label sample(1). Labeled records 3370(1) are produced during processing of sample(1) and used to label sample(2), and so on. Storage medium 3224 holds cumulative labeled records 3370(·). The count of cumulative labeled flow records 3380 after processing the last sample in the second labeler may be less than the count of the initial unlabeled flow records, the difference, referenced as residue 3385, may be further inspected for reasons other than labeling.

[0527] The training data available to the first labeler at the start of processing sample(k). k>0, includes cumulative labeled flow records {3320(0)+ . . . +3320(k−1)}, in addition to labeled flow records 3270(k−1). Likewise, the training data available to the second labeler at the start of processing sample(k). k>0, includes cumulative labeled flow records {3370(0)+ . . . +3320(k−1)}, in addition to labeled flow records 3220(k−1).

[0528] The switched labeled flow records from storage medium 3214 to updating process 3226 and from storage medium 3224 to updating process 3216 are indicated as 3250 and 3251, respectively. It is understood that in one embodiment, at the start of each sample, the cumulative labeled training records can be split with a portion of the training records utilized for one semi-supervised model which utilized packet size features and the remaining portion of the training records utilized for the second semi-supervised model which utilizes inter-arrival time features.

[0529] FIG. 34 illustrates composition 3400 of training data at start of labeling each sample of sampling stage according to a second co-training policy. The main difference between the two co-training policies is that the training data available to the first labeler at the start of processing sample(k). k>0, includes cumulative labeled flow records {3320(0)+ . . . +3320(k−1)}, in addition to cumulative labeled flow records {3370(0)+ . . . +3370(k−1)}. Likewise, the training data available to the second labeler at the start of processing sample(k). k>0, includes cumulative labeled flow records {3370(0)+ . . . +3320(k−1)}, in addition to cumulative labeled flow records {3320(0)+ . . . +3320(k−1)}. Thus, updated training data available to both labelers are the same, with the first labeler using packet-size parameters while the second labeler using time-domain packet-arrival patterns, including packet-inter-arrival time parameters. As well known, packet arrival patterns are analyzed as stochastic processes characterized using numerous parameters.

[0530] FIG. 35 illustrates an example 3500 of training-data age for a case of successive training epochs. A prescribed number of cumulative labeled/classified flow records 3520 over multiple epochs are held in a circular buffer (not illustrated) for use in extracting a training dataset of a manageable size.

[0531] FIG. 36 illustrates a first scheme 3600 of concurrent processes of training and real-time classification. During successive epochs of durations τ(0), τ(1), τ(2), . . . , flows received at Trafficwiz™ from clients are processed to determine respective classifications and records of successfully classified flows are communicated to concerned parities and also retained as additional training data for subsequent classification processes. Accumulated such records, reference 3640(·), during selected time periods each spanning two consecutive epochs, except for record 3640(1), are supplied to the training stage of Trafficwiz™ to update training data to be used for real-time classification during immediately succeeding epochs. For example, records 3640(2) of successfully classified flows between time instant t₀ and t₁ are supplied to the training stage at time t₁. The training stage updates the training data during a period of time T₂ to be used in Epoch(2) of duration τ(2).

[0532] FIG. 37 illustrates a second scheme 3700 of concurrent processes of training and real-time classification. Process 3720 acquires predetermined training data to be used for synthesizing a real-time classification model for Epoch(0).

[0533] During even-numbered epochs, starting with Epoch(0), concurrent operations 3740 comprising real-time operations 3741 and operations 3742 of training-data updating are performed. Processes 3741 receives flow-classification requests from clients and attempts to determine a respective class for each flow. Processes 3742 updates training data for an immediately succeeding epoch (an odd-numbered epoch) and synthesizes a real-time classification model to be used in processes 3761. Process 3750 switches to subsequent-epoch (an odd-numbered epoch) processing.

[0534] During odd-numbered epochs, starting with Epoch(1), concurrent operations 3760 comprising real-time operations 3761 and operations 3762 of training-data updating are performed. Processes 3761 receives flow-classification requests from clients and attempts to determine a respective class for each flow. Processes 3762 updates training data for an immediately succeeding epoch (an even-numbered epoch) and synthesizes a real-time classification model to be used in processes 3741. Process 3770 switches to subsequent-epoch (an even-numbered epoch) processing.

[0535] Continual activation 3780 of the loop of concurrent operations 3740 and concurrent operations 3760, necessitates continuous concurrent activation of the training component 110 and the real-time operational component 120 of Trafficwiz™ (FIG. 1).

[0536] FIG. 38 illustrates an arrangement 3800 of coupling processes of successive training epochs to update training flow records of successive epochs.

[0537] For a current epoch of index Π, Π≥0, training-phase 3820 performs processes of labeling and modeling. For Epoch(0), the processes are based on a set 2512 of records of unlabeled flows (FIG. 25). For subsequent epochs (Π>0), synthesis of the Trafficwiz™ modeler (the classification modeler) relies on a (growing) set of already labeled flow records 3810. A set 3850 of labeled flow records produced for epoch(Q) is supplied to process 3855 of age-filtering, applicable to Epochs succeeding the initial Epoch(0), where records of flows exceeding a prespecified age are excluded.

[0538] Real-time classification processes 3830 receive flow records 3825 from clients during Epoch(Q) and employs the Trafficwiz™ modeler (the classification modeler) synthesized in training phase 3820 to attempt to classify each received flow record. Process 3840 communicates classification results to concerned parties and concurrently provides the classification results to process 3860 which retains traffic-flow records and corresponding successful classifications to include in the set of training records of epoch(Q+1). Process 3870 of combines age-filtered flow records received from process 3855 and retained flow records from process 3860 to produce a set of training data 3880 for epoch(Q+1).

[0539] FIG. 39 illustrates exemplary outcome 3900 of the labeling stage for multiple training epochs. After processing Epoch(Q), Π≥0, the training data available to a subsequent epoch comprises labeled records 3920(Q) used at the start of Epoch(Q) and records 3940(Q) of successfully classified flows corresponding to classification requests received from clients. The combination 3950(Q) of labeled flow records after completion Epoch(Q) serves as an input to the training phase 3820 of a subsequent epoch. Where the size of combination 3950(Π) exceeds practical limits, a permissible number of labeled flow records may be extracted to supply as prior labeled records to a subsequent epoch.

[0540] FIG. 40 illustrates a growth pattern 4000 of training data. Initially, a set of unlabeled flow records, denoted Φ(0), is acquired and supplied to a labeler 760 which produces a set, denoted Λ(0), of labeled flow records. The number of labeled records of set Λ(0) may be less than the number of unlabeled flow records of set Φ(0). The difference being a “residue” that could not be decisively labeled. As described above, labeling a set of unlabeled flow records is performed for successive subsets of the unlabeled set as illustrated in FIG. 26.

[0541] The labeled set Λ(0) is used to synthesize a modeler 770 to be supplied to a real-time classifier 780 to classify individual flows received from clients 610 (FIG. 6). The period of time, which may vary significantly from a few hours to several days, during which a modeler based on a specific set of labeled flow records is used in the operational phase is herein referenced as an “epoch”. The duration of an epoch is a system parameter which may be based on a specified time interval or defined adaptively based on factors such as fluctuation of the volume of clients' traffic and/or availability of fresh unlabeled flow records.

[0542] Updating a modeler initiates a new epoch. Optionally, a set of flow records, denoted Θ(0), that are classified in the operational phase of the first epoch (referenced as Epoch(0) in FIG. 40) may be used as either labeled training data, to supplement the set of labeled flow records Λ(0), or as a set of new unlabeled flow records without using the already available classification. In the following, flow records classified in the operational phase are used to supplement already realized labeled flow records. The use of a combined labeled set Λ(0) and classified set Θ(0) to synthesise the modeler 770 leads to a new epoch, referenced as Epoch(1) in FIG. 40, where the modeler 770 is synthesized again based on the combined {Λ(0), Θ(0)} flow records. During Epoch(1), a set Θ(1) of clients' classified flow records is added to {Λ(0), Θ(0)} flow records and the modeler 770 is re-synthesized, for use in subsequent Epoch(2), based on labeled flow records {Λ(0), Θ(0), Θ(1)}.

[0543] During Epoch(2), a set Θ(2) of clients' classified flow records is created. Independently, a new set of unlabeled flow records, denoted Φ(1), is acquired (possibly automatically through network 220). Thus, at the start of Epoch(3), the training data comprises unlabeled set Φ(1), labeled set Λ(0), and classified sets {Θ(0), Θ(1), Θ(2)} denoted ΣΘ|2. The acquisition of Φ(1) triggers activation of labeler 760 to produce a corresponding set, Λ(1), of labeled records. In the example of FIG. 40, Λ(1) is produced in a training phase concurrently with the operational phase of Epoch(3). Growth of the training data continues in this fashion. Independently, fresh unlabeled flow records Φ(2) to Φ(6) are acquired during a time span of Epoch(4) to Epoch(24) and processed in labeler 760 to produce labeled sets Λ(0) to Λ(6). The training data at the start of Epoch(25) would comprise Φ(6), {Λ(0) to Λ(5)}, and {Θ(0) to Θ(24)}.

[0544] A growing training dataset may be pruned for processing considerations and to account for obsolescence of aged flow records. In the example of FIG. 40, cumulative counts 4020 of flow records labeled in the training phase and cumulative counts 4040 of clients' flow records classified in the operational phase may be subjected to pruning according to respective rules as discussed below with reference to FIG. 43. Preferably, labeled flow records and classified flow records are held in respective memory blocks since they may be pruned differently. Pruning may be performed using different methods. One method may be based on random sampling. Another method may be based on using a fast-clustering method, such as the “k-means” clustering method where a large number of flow records may be clustered into a smaller number of clusters with representative flow records selected based on proximity to centroids of the produced clusters. In this case, flow records further away from each cluster centroid would be discarded and sampled data from each cluster discarded for bias avoidance.

[0545] FIG. 41 indicates the mixture 4100 of training data comprising raw unlabeled flow record {Φ(0), Φ(1), . . . }, labeled flow records {Λ(0), Λ(1), . . . }, and classified flow records {Θ(0), Λ(1), . . . }. Unlabeled flow records 4110, denoted Φ(0) to Φ(6) and individually referenced as 4110(0) to 4110(6), respectively, are acquired from external sources. Processing flow records 4110 produces sets Λ(0) to Λ(6) of labeled flow records. Sets Θ(0) to Θ(24) of classified clients' flow records are produced in the operational phase of successive epochs, unlike sets 79 (0) to Λ(6) which are produced in the training phase when unlabelled flow records Φ(0) to Φ(6) are acquired. The mixture of labeled records Λ(0) to Λ(6) and classified records Θ(0) to Θ(24) at the start of successive epochs, starting with Epoch(1), are referenced as 4120(1) to 4120(24).

[0546] FIG. 42 illustrates modes 4200 of operation of Trafficwiz™ according to composition of training data. Process 4210 assembles training data at start of a current epoch. Process 4220 skips the labeling stage and proceeds to synthesize the classification modeler if the training data does not contain unlabeled flow records. Otherwise, process 4220 continues to activate the labeler in process 4230. Process 4240 synthesizes the classification modeler which is employed in process 4250 for real-time classification of clients' flow records. In the example of FIG. 40 and FIG. 41, process 4230 is activated at the start of Epoch(0), Epoch(3), . . . , and Epoch(24) as indicated in tabular form in FIG. 42.

[0547] FIG. 43 illustrates a process 4300 of pruning training data. Training data 4310 at start of an Epoch comprises cumulative labeled flow records 4320, denoted Σ(κ|J), J≥0, and cumulative classified clients' flow records 4340, denoted Σ(Θ|Q}, Q≥0.

[0548] A first pruning filter (Λ-pruning filter) 4330 selects labeled flow records to retain based on a first criterion. A second pruning filter (Θ-pruning filter) 4350 selects classified flow records to retain based on a second criterion. Both the first and second criteria may select flow records according to age, number of accumulated records, and a weighting factor relevant to the type of flow record (belonging to 4320 or 4340). The weighting factor may be a design parameter.

[0549] The Classifier 780 combines the trained AI model created using the Modeler 770, with a packet capture and feature extraction engine (not shown), creating an efficient application for classifying encrypted network traffic. The Classifier 780 may be operated in two modes. In an “Offline mode”, unlabeled encrypted flows are classified. In an operational “live mode”, live network flows, which maybe encrypted, are captured and classified.

[0550] In one embodiment, a hierarchical classifier is employed, with real-time runtime performance and improves classification results. With a hierarchical classifier, several Machine-Learning models are used in succession to determine labels ultimately assigned to flows. The Machine-Learning models form a tree structure, where, aside from the root node, the input to one classifier is based on the output of another.

[0551] While the hierarchical classifier may comprise an arbitrary structure, with as many classification nodes as required to serve the intended purpose, one embodiment of the present invention has only two levels of classification. The first level is a binary classifier, splitting traffic flows into those containing payload and those without the payload. The non-payload flows are filtered out, passing on only payload carrying flows. The second level comprises a multi-class model, designed to infer labels for payload carrying flows. Traffic types can be specified in a configuration file. Traffic that does not belong to specified traffic types will be classified into the “Unknown” category. In practice, any suitable ML/DL model may be used for any node in the tree of hierarchical classifiers, allowing each to be trained independently, optimized for a specific purpose.

[0552] Non-payload flows, removed by the first level of the hierarchical AI model, include those which do not contain user content, such as application protocol handshake and signaling. It may also include some ancillary flows which are generated by, but do not pertain to, the main application; such content could confuse training models. The result of separating payload and non-payload carrying flows is a clean set of traffic flows, improving the training and performance of the second level model. It also improves the runtime performance of the classifier operating in live environments, since non-payload carrying flows are rejected early in the pipeline.

[0553] Using the ROC (Receiver Operating Characteristics) analysis, a preferred threshold is one that favors one type of error over the other; reducing false positives at the cost of true positives, or vice versa. It is useful to tune the model so that no payload data is excluded during the classification/prediction. The objective is to classify payload data with some latitude for false classification of non-payload traffic as payload traffic. The performance of the first level of classifier can be tuned by selecting a threshold on the ROC curve. The value is selected using hyper-parameter search procedure. It is understood that there are alternate methods which can be utilized to tune the classification model so that it achieves the trade off between accuracy and real-time performance.

[0554] FIG. 44 illustrates an implementation 4400 of the AI model and classifier 780. A Receiver-Operating-Characteristics (ROC) module 4420 is employed to tune the model so that it achieves class separation when predicting an unlabeled flow record 4410 as belonging to a non-payload class 4440 or a payload-carrying flow class 4450. A multi-class model 4460, configured to infer labels for flows, is subsequently used to classify the payload-carrying flow data 4450.

[0555] The proposed classifier 780 can support any ML/DL algorithm for model creation during Step 1 (box 4420) and Step 2 (box 4460). For example, the Random Forest algorithm may be utilized for the Step 1 model building while the XGBoost algorithm may be used for model creation during the Step 2 model building.

[0556] Binary classifiers, such as the Step 1 ML model, generally produce a score that is compared against a predetermined threshold to determine whether the result belongs to one class or the other. The threshold is chosen to give equal weight to both classes, but results can often be improved by adjusting the threshold. As mentioned above, one method of choosing an optimal threshold is by an analysis of the Receiver Operating Characteristic (ROC) curve. The ROC curve captures performance of a model by graphing the True Positive Rate (TPR) against False Positive Rate (FPR) for different threshold values.

[0557] The process for generating a ROC curve is as follows:

• • [0558] Using a trained model, generate scores for a validation dataset.
  • [0559] For a selected number of threshold values, determine labels for entries in the validation dataset using the scores produced in the first step. For example, one may evaluate 10 or 100 values between 0 and 1 when using probability scores.
  • [0560] For each threshold value, calculate TPR and FPR, given ground truth labels and predictions made using that threshold
  • [0561] Present the set of TPR vs FPR values graphically.

[0562] The optimal choice of threshold to maximize TPR while minimizing FPR is a threshold which produces a point closest to the upper-left corner of the ROC curve, as illustrated for example in FIG. 44A by the ROC diagram 4400A for selecting an operational threshold for the binary classification.

Working Example for Developing a Model which Distinguishes “Payload Data” v “Non-Payload Data”

[0563] As mentioned above, a Payload flow is an application specific flow which carries user data for the application.

[0564] Non-payload flow are flows which carry limited information, or flows which are not generated by application related activity. Non-payload flows include examples such as signaling, protocol handshake, advertisements (e.g. flows generated by DoubleClick and AMP project), control flows (e.g. flows generated by Google API, Amazon AWS, Meta Platform CDN) etc.

[0565] Non-payload flows contain both long flows and short flows. If the flow's bidirectional packet count is less than 10 (forward_pkt_count+backward_pkt_count<10), the flow is considered a short flow. Short flows are categorized as Non-payload. Long flows ((forward_pkt_count+backward_pkt_count>=10), can be Payload carrying flow or a Non-payload flow. Identification of long Non-payload flows is achieved using the 1st Step ML model outlined earlier.

[0566] In this working example, predicting a flow as belonging to Payload vs Non-payload is achieved as follows.

• • [0567] 1. The ML Model is created using training data. The training data includes apriori labeled flows with Payload and Non-payload labels in the training dataset. The training data flows are labeled as one of Audio Chat, Video Chat, Text Chat, Audio Stream, Video Stream and P2P for the “Payload” training data, with other flows labeled as “Non-payload” training data.
  • [0568] 2. Network Traffic Features utilized include
    • [0569] Packet length min, max, mean and standard deviation for both forward and backward directions;
    • [0570] Inter packet arrival time min, max, mean and standard deviation for both forward and backward directions;
    • [0571] Packet count for both forward and backward directions;
    • [0572] Total packet length for both forward and backward directions;
    • [0573] Total payload size for both forward and backward directions;
    • [0574] Number of payload packets for both forward and backward directions;
    • [0575] First payload packet size for both forward and backward directions;
    • [0576] Flow duration; etc.
    • [0577] The working example includes 27 network traffic features.
  • [0578] 3. Algorithm XGB is applied, as cited in the Information Disclosure Statement, see Tianqi Chen and Carlos Guestrin “XGBoost: A Scalable Tree Boosting System”.
  • [0579] 4. ROC curve and determining the point on ROC curve during training

[0580] In network traffic analytics, payload carrying flows are more important than Non-payload flows. In a binary classifier, if one class is more important rather than the other, the probability or the ‘roc_threshold’ value can be calibrated to reduce errors and to classify one class over the other. The binary classifier predicting Payload-Nonpayload is tuned by selecting ‘roc_threshold’ from the ROC (Receiver Operating Characteristic) curve. ROC provides a useful graphical representation to evaluate the performance of a binary classifier and tune the AI model. The ROC curve illustrates the trade-off between ‘True Positive Rate’ (tpr) and ‘False Positive Rate’ (fpr) for a predictive model using different probability thresholds or referred to as ‘roc_threshold’. This threshold value can be obtained from the ROC curve, or the best threshold value can also be calculated using Youden's J statistic: j=tpr−fpr. Please refer to the paper by W. J. Youden “Index for Rating Diagnostic Tests” cited in the Information Disclosure Statement.

[0581] The value of ‘roc_threshold’ ranges from 0.0 to 0.5. For example, the value 0.5 for this parameter represents that the first level classifier would not admit any flow into the second level for further analysis, if the flow prediction probability for the non-payload flow class is >=50%.

[0582] In practice, the model is run multiple times (5 times is a reasonable example) on train-test split data, then the mean value of the ROC threshold may be assigned as the optimal ROC threshold. The train-test ratio starts with about 50%-50% and incrementally increases the training portion of the data by 10% until the ratio reaches around 90%-10%.

• • [0583] 5. At the operation stage

[0584] During the operation stage, the aforementioned network features are used by the HXGB classifier for its model. The first level XGB classifier will admit flows with prediction probability greater than the ROC threshold into the second level XGB traffic type classifier for further classification, while classifying the remaining flows as Non-payload.

Examples of Training-Dataset-Growth Control

[0585] FIG. 45 illustrates an example 4500 of growth of a training dataset with unconstrained inclusion of training data acquired from the real-time classification stages over a period of 40 epochs of predetermined durations. The size 4510 of the training flow records at the beginning of each epoch succeeding Epoch(0) equals the sum of the size 4520 of an initial training dataset produced in Epoch(0) and the size 4530 of training data acquired from the real-time classification processes during preceding epochs.

[0586] FIG. 46 illustrates an example 4600 of growth of a training dataset with constrained inclusion of training data acquired from the real-time classification stages according to a first policy of age filtering of training data. The size 4610 of the training flow records at the beginning of each epoch succeeding Epoch(0) equals the sum of the size 4620 of a subset of the initial training dataset 4520, selected according predefined rules, and the size 4630 of age-filtered training data extracted from the real-time classification stages of preceding epochs.

[0587] FIG. 47 illustrates an example 4700 of growth of a training dataset with constrained inclusion of training data acquired from the real-time classification stages according to a second policy of age filtering of training data. The size 4710 of the training flow records at the beginning of each epoch succeeding Epoch(0) equals the sum of the size 4520 of the initial training dataset and the size 4730 of age-filtered training data extracted from the real-time classification stages of preceding epochs.

The Modeler 770

[0588] Modeler 770 enables development, experimentation and validation of ML models for encrypted traffic classification. Modeler 770 is configured to support and create a ML pipeline with data pre-processing, transformation and a suitable ML algorithm. Feature selection, hyper-parameter tuning and various data transformations are utilized. Multiple iterations of ML models are built and iteratively evaluated to yield improved classification results. This process includes the analysis of results followed by modification of the pipeline or parameters and re-testing.

[0589] The Modeler provides a flexible platform to build and experiment with ML Models. The ability to select various pipeline functions and parameters are key to its design. The user creates an ML model using pipeline workflows, algorithm. and various parameters set in a configuration file.

[0590] ML pipeline functionalities and algorithms are based on existing libraries (e.g., scikit-learn, TensorFlow) or homegrown custom algorithms. Users are given an ability to experiment and build models using various machine learning pipelines and workflows. The final model may be exported in Open Neural Network Exchange (ONNX) format. The ONNX format allows the user to easily transition between various machine learning frameworks and tools and to leverage the capabilities of different platforms for the same ML models.

[0591] The Modeler 770 is configured to:

• • [0592] (d) Extract bi-directional flows and features for network traffic stored in unlabeled traffic-flows files. A set of 135 features (statistics) are computed for each network traffic flow. A flow is defined by a 5-tuple—Source IP address, Destination IP address, Source Port, Destination port and Protocol. A flow contains one or multiple packets. Flows in a network are bi-directional. Packets from the client to the server are part of forward flows (upstream flows). Packets from the server to the client are part of backward flows (downstream flows). Manipulate data and perform pre-processing of feature statistics computed in the earlier step. This includes data cleansing, creating balanced datasets, outlier detection and data augmentation.
  • [0593] (e) Create machine learning models based on user configurable ML pipelines. ML pipeline includes mechanism for data transformation, feature selection and model creation.
  • [0594] (f) Train an ML model and evaluate against test dataset and training dataset. Perform hyper-parameter optimization to identify model with best performance.
  • [0595] (g) Export the trained model to ONNX format for use in TrafficWiz Classifier for classification in a run-time environment.
  • [0596] (h) Create ML models to operate offline or in real-time. Offline operation is not performance-sensitive and the steps outlined above can be followed. However, for real-time operation, the ML models created above require additional fine tuning as there are time constraints on how long the Classifier 780 takes to classify a traffic flow in real-time. To enable a fast decision by the Classifier, the ML model built by Modeler 770 takes into account two key criteria:
    • [0597] (i) a decision needs to be made about the traffic class of the flow within the first 30 packets, or so. The system cannot wait for a large number (1000, for example) of packets in a flow, in the case of a Video Stream, for example. The system ensures the tradeoff and identifies that approximately 30 packets is sufficient to achieve the desired accuracy while meeting the performance constraints.
    • [0598] (ii) Computation of all 135 statistics is not practical in performance driven classification systems. As a result, the system of the present invention seeks to find the minimum number of statistics that can be computed without sacrificing accuracy. A sub-set of about 35 features (from about 135 features) achieves real-time classification performance while retaining accuracy to a desired level.

[0599] As outlined above, the ML pipeline supported by Modeler 770 includes a Feature Selection method using a genetic algorithm. The data features used to train the machine learning models have a significant influence on their performance. The irrelevant or partially relevant features can negatively impact the model performance. Feature selection is the process of reducing the number of input variables when developing a predictive model. The feature selection helps to reduce the computational cost of modeling and may improve the performance of the model.

[0600] After the feature computation, a sub-set of features are chosen for most effective prediction results. Thus, the 135 computed features are not required for best prediction for the classifier 780. A properly selected sub-set of features is sufficient. A feature selection algorithm for network flow feature selection is developed based on the use of Genetic Algorithms (GA), as will be described in more detail with regard to FIG. 49 below.

[0601] Genetic Algorithms are just one of a family of algorithms that aim to find optimal solutions to optimization problems. At a high level they share a general optimization process: candidate solutions are generated, evaluated, and ranked over many iterations, leading to high value or optimal results. Inspired by nature, candidate solutions in GAs are represented as strings of chromosomes and undergo genetic operations to produce subsequent generations of candidates. Two such genetic operations are mutation, which alters a single gene, and reproduction through crossover, which combines the genetic sequences of two parents to create two offspring.

[0602] The problem of flow feature selection is to choose the subset of features that provide the best result on the desired task. Candidate solutions are encoded as sequences of Boolean variables, indicating the presence or absence of a particular feature. In a bit string 0 and 1 will indicate inclusion or exclusion of one feature. Initially, a pool of several-hundred candidate solutions is created by random selection. Each iteration of the algorithm ranks the current set of candidate solutions and generates a new set through genetic operations. In this case, the ranking process is to train and test a classification model using the subset of features specified by the candidate solution, using the F1 score of a test set as an evaluation metric. Creation of the next generation involves repeatedly selecting two parent candidate solutions from the ranked list and applying mutation and crossover to produce new candidate solutions; higher scoring candidates are chosen to be parents with proportionately higher probability. The next iteration begins once enough solutions have been generated, and continues until the solution converges or for a fixed number of iterations.

[0603] FIG. 48 illustrates a machine-learning pipeline 4800 for synthesizing modeler 770. A storage medium 4805 holds labeled flow records received from labeler 760. A set of pre-processing modules 4835 includes a data-cleansing module 4810, a data imputation module 4820, and a module 4830 for removing flow records considered to be outliers. The set of pre-processing modules 4835 may not be needed if data cleansing, imputation, and filtering to remove outliers are performed in the training phase. A set of feature-related modules 4875 includes a feature-computation module 4850, a feature-extraction module 4860, and a feature-selection module 4870.

[0604] A data-transformation module 4840 prepares the pre-processed data, received from modules 4835, for processing in the feature-related modules. A model-tuning module 4880 prepares the model for use in real-time classification of clients' unlabeled flow data. A memory device 4890 holding an encoded modeler.

[0605] FIG. 49 illustrates a method 4900 for network feature selection using a genetic algorithm, and creating a machine learning classification model for the Modeler 770 based on the selected network features, for use in the Classifier 780.

[0606] Upon Start of the method 4900, N network traffic features characterizing the encrypted network traffic are collected/obtained, and an input (initial) set of candidate solutions is formed, which is a set of randomly selected traffic features, from the N network traffic features (box 4910). The input set of candidate solutions is feed in the genetic algorithm, performing crossover and mutations operations on the input set of candidate solutions (box 4920), and outputting an output set of candidate solutions, which is an output set of selected network features (arrow from box 4920 to box 4930).

[0607] Next, respective classification machine learning models for classification of the encrypted network traffic are created, corresponding to the output set of selected network features (box 4930). The classification models are trained and evaluated, where, for each model, an accuracy of classification results by each model is determined and compared with ground truth network traffic classification results (box 4940). Provided exit criteria for terminating the method have not been met (Exit No from box 4950), the output set of candidate solutions is assigned as the input set of candidate solutions, and the steps (b) to (d) are repeated (arrow from box 4950 to box 4920).

[0608] Provided exit criteria for terminating the method have been met (exit Yes from box 4950), choosing such machine learning model based on selected output network features that provides the highest accuracy, or highest ranking (box 4960), followed by termination of the method 4900. Exit criteria in box 4950 may include a predetermined number of iterations, or a predetermined accuracy of the classification model with the highest ranking. If the accuracy of the classification model with the highest ranking is still below a predetermined required accuracy, exit criteria may be updated for example by increasing the number of iterations, or increasing the total number of N network features. For example, we had about 30 network traffic features when we started our experiments. Then we have added more than 100 network traffic features after conducting a research for over 3 months. AI ternatively, we may have over 100 network traffic features available at the beginning, but decided to use only a subset of the features for reducing computational time, which number is further reduced by the genetic algorithm.

[0609] A total number of collected network traffic features may be fairly large, for example around or below a few hundred features, for example below or equal to N=500 features. Some features such as byte distribution alone may contain 256 features. AI ternatively, N may be chosen equal to about 135; or N may be in a range from about N=70 to about N=150; or N may be in a range from about N=30 to about N=135.

[0610] The reduced number M of network traffic features after the genetic algorithm has been from about 10 to about 30 features in our experiments, with about 30 features being used for 100 Gbps network traffic, and about 10 features being used for Tb/s network traffic.

[0611] Another factor is a number of packets monitored to compute the statistics/feature in each flow. For currently available computational processing powers, to achieve 1 Tbit/second, not only M should be around 10 features, but the number of packets used to compute the features should be limited to 15-40 packets. Using less than 15 packets could provide speed but reduces accuracy. It is understood that the number M of selected features, and the number of packets may be increased in the future as the computational power of processors increases. In general, the number of M selected network features and the number of monitored packets for each network feature are selected experimentally, depending on the data flow rate of the network traffic: for the same processing power, the higher is the data flow rate of the network traffic, the lower is the number of M selected features and the lower is the number of monitored packets. The lowest number of M selected features and the monitored packets is determined by a low threshold of an acceptable accuracy.

[0612] FIG. 50 illustrates a closed-loop flow-classification arrangement 5000 of a labeler 760, a modeler 770, and a classifier 780. The labeler receives external training data 5010 in addition to data relevant to classified clients-defined flows which is feedback to the labeler to be used as training data. Clients' classification requests 5020 are processed to produce classified-flow data 5030 sent to clients. Classified-flow data 5040 are sent to the labeler for use as additional training data.

[0613] FIG. 51 illustrates a method 5100 of concurrent activation of labeler 760, synthesis of modeler 770, and real-time classification at classifier 780 as illustrated in FIG. 40. Definitions of the notations Φ, Λ, Θ, ΣΛ, and ΣΘ are indicated in the figure. Process 5110 initializes variables. Process 5120 receives fresh training data. Process 5130 branches to process 5140 to activate the labeler if any fresh training data has not been processed. Otherwise, process 5160 determines if there is a sufficient number of labeled flows to synthesize the modeler in process 5170. Process 5175 continuously receive clients' classification requests which are processed in the classifier (reference 5180). Process 5150 and process 5190 update results.

The following notation is used in FIG. 51:

• • [0614] Φ: Fresh training flow data
  • [0615] Λ: Labeled flow data of Φ
  • [0616] Θ: Classified flow data derived from clients' classification requests
  • [0617] ΣΛ: Cumulative value of Λ
  • [0618] ΣΘ: Cumulative value of Θ

[0619] Process 5110 initializes ΣΛ and ΣΘ. Process 5120 receives fresh training data, Φ, at different instants of time which may be irregularly spread in the time domain. A received collection of training-flows records is preferably, or inevitably, handled as successive samples of limited sizes (limited number of records) each, as illustrated in FIG. 26. Thus, upon receiving such a collection, process 5130 determines whether the entire collection is presented to the labeler 760. If there are pending training training-flow records, the labeler is activated in process 5140 to produce more labeled flow records Λ and process 5150 updates ΣΛ according to results of process 5140. If the entire collection has been processed in the labeler 760, process 2160 determines whether a sufficient number of labeled records, at least equal to a predefined value Σ_min, has been reached.

[0620] If ΣΛ is less than, Σ_min, process 5120 is revisited to wait for receiving more training data. Otherwise, process 5170 is activated to synthesize an updated modeler 770. The updated modeler is used in process 5180 to classify clients-specified flow records extracts from received classification requests of process 5175. Upon updating ΣΘ in process 5190, process 5170 is revisited. Meanwhile, independent process 5120 and process 5175 continue ad infinitum to receive training data and clients' requests, respectively.

[0621] In general, Encrypted network traffic flows are analyzed using multiple techniques including:

• • [0622] (1) Clustering of flow data using unsupervised machine learning clustering algorithm Density-based Spatial Clustering of Applications with Noise (DBSCAN);
  • [0623] (2) Mapping IP address to DNS name to obtain results from reverse-DNS (IP address to DNS name mapping is used to infer flow network traffic type);
  • [0624] (3) Extracting Server Name Indicators from TLS traffic exchange (also used to infer flow network traffic type);
  • [0625] (4) Obtaining results from Deep Packet Inspection software (also used to infer flow network traffic type);
  • [0626] (5) Using previously trained supervised ML model, and
  • [0627] (6) Inspecting network traffic characteristics including for example packet sizes, total bytes, number of packets, flow duration, bit rate in both forward and backward directions.

[0628] Applying the above techniques facilitates suggesting a label for a traffic flow.

[0629] The results may be collectively used to infer a label for a network traffic flow.

[0630] A proposed heuristic attempts to identify a particular class of traffic by analyzing its traffic characteristics using a set of rules. Table 1 presents results for six traffic classes (Audio Chat, Video Stream, Audio Stream, Text Chat, Video Chat and Peer-to-Peer).

[0631]

TABLE 1
Results of applying a labeling heuristic for common payload flows.
Traffic	Payload	Filter
Type	Character	Applied	Feature Criteria Reference
Audio	Both	Total	The ‘mean packet size’ forward and backward are
Chat	directions	time ≥2	comparable - between 100 and 200 bytes.
			The flow ‘packet count’ forward and backward are
			approximately the same for mobile and wireline data. They
			are comparable and usually more than 1000. The ‘flow
			payload size’ forward and backward are large (e.g. flow
			payload size in one direction should be more than 0.1 MB)
Video	Backward	Total	The ‘mean packet size’ forward is less than 100, and ‘mean
Stream	direction	time ≥2	packet size’ backward is more than 1000.
			The ‘flow payload size’ backward is significantly larger than
			forward direction.
			The rate is usually more than 400 kbps.
Audio	Backward	Total	The ‘mean packet size’ forward is less than 100, and ‘mean
Stream	direction	time ≥2	packet size’ backward is more than 1000.
			The flow duration is usually more than 10 s.
			The rate is more than 1.6 kbps but less than 150 kbps.
			If rate is greater than 150 kbps but the flow duration is less
			than 30 s then these flows are still strong candidate of payload
			data.
Text	Both	Total	The ‘mean packet size’ forward and backward should be
Chat	directions	time ≥2	around 100 and comparable.
			The flow ‘packet count’ forward and backward are
			comparable and more than 10.
			The flow duration is usually more than 5 s.
			The rate can be very small and usually less than 10 kbps.
Video	Both	Total	The ‘mean packet size’ forward and backward should be
Chat	directions	time ≥2	similar and more than 500.
			The flow ‘packet count’ forward and backward is
			comparable.
			The flow duration is usually more than 10 s.
			The rate is usually more than 1000 kbps.
Peer-	Both	Total	The rate is usually more than 100 kbps.
to-	directions	time ≥5	One direction of ‘mean packet size’ is usually over 1000, the
Peer			opposite direction is usually less or around 100
			The flow duration is usually more than 10 s.

[0632] It has been found experimentally that adding a small fraction of network flows from the real-time encrypted traffic from the deployment site to the training data set, improves the performance of the Labeler 760. The fraction of the network flows needs to be representative of all types of all real-time encrypted traffic flows and substantially smaller than the training data set. We refer to this technique as “salting”.

[0633] Consider two datasets, dataset-A and dataset-B, which are captured separately from network traffic at different times and different locations. When a model is trained on dataset (A) and tested on dataset (B), the performance of the classification model is lower than when the same model is trained on one portion of dataset-A and tested on another portion of dataset-A.

[0634] In a salting process, a small portion (0.02 to 0.05) of dataset-B, covering all the traffic types, is added to the training dataset-A on which the model is to be trained. When the model is trained on the new augmented dataset, comprising dataset-A plus a relatively small number of flows of dataset-B, the model also learns the temporal and spatial characteristics of dataset-B. As a result, the trained model can predict the traffic classes with higher accuracy.

[0635] In a test case, a dataset containing 5 traffic classes was used to train the model; the traffic classes being audio chat, audio stream, peer-to-peer, text chat, and video stream. The training dataset-A was captured in year 2020 on a wireline network. Dataset-B also consists of same set of traffic classes but was captured in the year 2019 on a wireline network.

[0636] In test-case-1 (Table-A), a model using a Random Forest classifier is trained on dataset-A and tested on dataset B, and the overall performance (F1-Score) of the model was determined as 67.6%. However, when a small portion of dataset-B (0.02 to 0.05) is added to training dataset-A and tested on the remaining part of dataset-B, a significant increase in overall F-score of the model is observed (test-case-2, test-case-3). Using the same methodology, in test-case-4, test-case-5, and text case-6, similar results were observed but with a different model using an XGBoost classifier.

[0637]

TABLE A
Test cases based on dataset A and B
Test				F1-Score
case	Training Dataset	Test Dataset	Model	(%)
1	A	B	Random	67.6
2	A + 2% flows of B	98% flows of B	Forest	85.79
3	A + 5% flows of B	95% flows of B	Classifier	87.19
4	A	B	XGBoost	58.65
5	A + 2% flows of B	98% flows of B	Classifier	84.14
6	A + 5% flows of B	95% flows of B		85.83

[0638] In another example, the same experiment was repeated but with different test dataset (Table-2). In this experiment a dataset-C was used as a test dataset, which was captured in year 2020 using a 5G network.

[0639]

TABLE B
Test cases based on dataset A and C
Test				F1-Score
case	Training Dataset	Test Dataset	Model	(%)
1	A	C	Random	73.11
2	A + 2% flows of C	98% flows of C	Forest	85.12
3	A + 5% flows of C	95% flows of C	Classifier	86.55
4	A	B	XGBoost	69.65
5	A + 2% flows of C	98% flows of C	Classifier	85.09
6	A + 5% flows of C	95% flows of C		86.13

[0640] In test-Case-1 (Table-B), the model using a random forest classifier is trained on dataset A and tested on dataset C, and the over all performance (F1-Score) of the model is 73.11%. However, when a small portion of dataset-C is added to training dataset A, and tested on the remaining dataset, a significant increase in overall F1-score of the model (Test case2, testcase-3). Using the same methodology, in test case-4, test case-5, and test case-6, similar results were observed but with a different model using XGBoost classifier.

[0641] It should be noted that methods and systems of the embodiments of the invention and data sets described above are not, in any sense, abstract or intangible. Instead, the data is necessarily presented in a digital form and stored in a physical data-storage computer-readable medium, such as an electronic memory, mass-storage device, or other physical, tangible, data-storage device and medium. It should also be noted that the currently described data-processing and data-storage methods cannot be carried out manually by a human analyst, because of the complexity and vast numbers of intermediate results generated for processing and analysis of even quite modest amounts of data. Instead, the methods described herein are necessarily carried out by electronic computing systems having processors on electronically or magnetically stored data, with the results of the data processing and data analysis digitally stored in one or more tangible, physical, data-storage devices and media.

[0642] Methods and systems of the present invention have tangible and practical advantages, providing more expedient and more reliable identification of network topology, and classification and identification of devices in networks.

[0643] Systems and apparatus of the embodiments of the invention may be implemented as any of a variety of suitable circuitry, such as one or more microprocessors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), discrete logic, software, hardware, firmware or any combinations thereof. When modules of the systems of the embodiments of the invention are implemented partially or entirely in software, the modules contain a memory device for storing software instructions in a suitable, non-transitory computer-readable storage medium, and software instructions are executed in hardware using one or more processors to perform the techniques of this disclosure.

[0644] While the above description provides examples of the embodiments, it will be appreciated that some features and/or functions of the described embodiments are susceptible to modification without departing from the spirit and principles of operation of the described embodiments. Accordingly, what has been described above has been intended to be illustrative of the invention and non-limiting and it will be understood by persons skilled in the art that other variants and modifications may be made without departing from the scope of the invention as defined in the claims appended hereto.