Encrypted Flow Classification Method and System for Multi-View Consistency and Progressive Observation Control

The encrypted flow classification method based on multi-view consistency and progressive observation control solves the problems of insufficient generalization and high edge deployment overhead in existing encrypted traffic classification technologies, and achieves efficient and reliable encrypted flow classification in real network environments.

CN122951877BActive Publication Date: 2026-11-17SOUTHWEAT UNIV OF SCI & TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611432879.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-16
Publication Date
2026-11-17
Estimated Expiration
2046-09-16

AI Technical Summary

Technical Problem

Existing encrypted traffic classification methods do not generalize well in real network environments, are prone to misjudgment, have difficulty identifying unknown traffic, and have high overhead for edge deployment, failing to meet the requirements of low latency and low computing power.

Method used

A cryptographic flow classification method based on multi-view consistency and progressive observation control is adopted. By aggregating cryptographic flows to form flow-level analysis objects, multiple view sketches are generated. The credible decision score is calculated by combining candidate category confidence margin, view consistency, stability between adjacent observation budgets, and environmental calibration items, so as to realize progressive observation and escalation of actions.

Benefits of technology

It significantly reduces the risk of misclassification in scenarios with insufficient evidence, shortens decision latency, reduces edge computing power consumption, improves the system's adaptability to environmental changes and unknown traffic, and supports the efficient deployment of edge nodes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122951877B_ABST
    Figure CN122951877B_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for classifying encrypted streams with multi-view consistency and progressive observation control, relating to the field of network security technology. The method includes collecting and aggregating encrypted streams, setting multiple incremental observation budgets, generating at least two view sketches under the current observation budget and classifying them, obtaining the category probability distribution and embedding representation of each view sketch, fusing them to obtain the current candidate category, calculating the confidence decision score under the current observation budget based on the multi-view classification results, and considering whether the observation budget has reached its maximum, selecting either to output the classification result and stop observation, update to the next adjacent observation budget and continue observation, or output a low-confidence result, an unknown result, or a risky result. The higher-level analysis node further determines the outcome and executes corresponding actions based on the final judgment. This invention can reduce the risk of misclassification under insufficient evidence conditions, shorten the average decision latency, reduce edge computing power consumption, and improve the system's adaptability to environmental changes and complex traffic.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of network security technology, and more particularly to the field of encrypted traffic analysis and intelligent classification technology; specifically, it relates to a multi-view... Figure 1 A method and system for classifying encrypted streams with consistency and progressive observation control. Background Technology

[0002] With the widespread application of protocols such as HTTPS (Hypertext Transfer Security), TLS (Transport Layer Security), and QUIC (Quick Internet Connection with UDP), encrypted traffic classification has become a fundamental capability in network security protection, service identification, resource scheduling, and anomaly detection. In recent years, related research has gradually expanded from closed-world and static distribution scenarios to scenarios that are closer to real-world networks, such as environment drift, non-independent and identically distributed networks, open set identification, and line-rate deployment.

[0003] Most existing encrypted traffic classification methods make a one-time classification decision based directly on fixed-length prefixes, complete session features, or a single view representation. Although they can achieve high accuracy on ideal datasets, they still have the following shortcomings in real networks:

[0004] (i) Changes in network environment, fluctuations in link quality, differences in terminals and evolution of services can cause shifts in feature distribution, making fixed models prone to misjudgment, i.e., insufficient generalization under environmental changes;

[0005] (ii) Existing methods tend to force the output of a known category when there is insufficient evidence, which is not conducive to the robust identification of unknown traffic or low-reliability traffic;

[0006] (iii) Some methods rely on long observation windows or heavy models, making it difficult to simultaneously meet the requirements of low latency, low computing power and edge deployment. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-view Figure 1 A cryptographic flow classification method and system based on consistency and progressive observation control is proposed to address the problems in existing technologies, such as misclassification of unknown traffic, insufficient generalization under environmental changes, premature hard judgment, and high overhead of edge deployment.

[0008] The present invention solves the above problems through the following technical solution:

[0009] A multi-view Figure 1 Encrypted flow classification methods based on consistency and asymptotic observation control include:

[0010] Step S1: Aggregate the collected encrypted streams to be classified to form stream-level analysis objects; set the observation budget sequence, which consists of multiple incremental observation budgets, and select the first observation budget to proceed to the next step;

[0011] Step S2: Extract the prefix message subsequence of the flow level analysis object under the current observation budget;

[0012] Step S3: Generate at least two view sketches from different feature dimensions based on the prefix message subsequence. The view sketches include, but are not limited to, length-direction sketches, burst sketches, time-normalized sketches, state transition sketches, and packet loss-resistant sketches, so as to characterize the behavior pattern of encrypted traffic from multiple perspectives. Preferably, the prefix message subsequence is preprocessed beforehand. The preprocessing includes one or more of the following: direction marking, time alignment, anomaly filtering, or length discretization.

[0013] Step S4: Input the view sketches into the corresponding classification models to obtain the category probability distribution and embedding representation of each view sketch; obtain the fusion category probability distribution under the current observation budget based on the category probability distribution of each view sketch, and determine the candidate category with the highest probability in the fusion category probability distribution as the current candidate category;

[0014] Step S5: Based on the category probability distribution and embedding representation, calculate the credible decision score under the current observation budget. The credible decision score is composed of the candidate category confidence margin, the apparent confidence margin, and the apparent confidence margin. Figure 1 The stability is determined jointly by consistency, stability between adjacent observation budgets, environmental calibration terms, and instability penalty terms;

[0015] Step S6, Judgment:

[0016] When the credible decision score is greater than or equal to the acceptance threshold corresponding to the current candidate category under the current observation budget, the classification result is output and observation is stopped.

[0017] When the credible decision score is less than the acceptance threshold and the maximum observation budget has not been reached, the next observation budget adjacent to the current observation budget in the observation budget sequence is selected, and the process returns to step S2 to extract the extended prefix message subsequence from the starting position of the same flow level analysis object.

[0018] If the confidence decision score is still less than the acceptance threshold when the maximum observation budget is reached, a low confidence result, an unknown result, or a risk result is output, and an upgrade handling package is sent to the superior analysis node for upgrade handling. The system also receives the final judgment result returned by the superior analysis node and executes the corresponding handling action.

[0019] As a further improvement of the present invention, the observation budget is characterized by at least one of the following: the number of prefix messages, the prefix time window length, the number of message direction changes, and the number of burst units. For example, an incremental observation budget sequence is set for the budget representation method used. Among them, observation budget This represents the upper limit of observations accumulated from the starting position of the flow level analysis object to the end point of the t-th level observation, where 1≤t≤n, n is the number of observation budgets, and t is the number of the observation budget.

[0020] As a further improvement of the present invention, the credible decision score is determined by the following formula:

[0021] ;

[0022] in, The score represents the credible decision score. Confidence margin for candidate categories, For vision Figure 1 Devotion, For the stability between adjacent observation budgets, For environmental calibration items, As an unstable penalty item, Preset weights for the confidence margin of candidate categories; For vision Figure 1 Preset weights for consistency; Preset weights for stability between adjacent observation budgets; Preset weights for environmental calibration items; The preset weights are for the instability penalty term. Among them, the candidate class confidence margin... ,See Figure 1 Desire Stability between adjacent observation budgets and instability penalty items Calculated from current and historical classification outputs; Environmental calibration item The confidence margin of the candidate category is obtained by statistically analyzing the multi-network environment calibration dataset according to the "network environment-candidate category-observation budget"; ,See Figure 1 Desire Stability between adjacent observation budgets and instability penalty items Environmental calibration items All values ​​were normalized before weighted calculation. The intervals α, β, γ, η, and δ are determined during the offline calibration phase through parameter search or constraint optimization. Simultaneously, the corresponding acceptance threshold is learned for each candidate weight group. Under a preset error acceptance rate constraint, the weight combination with the fewest average decision rounds is selected. For example, for each candidate weight combination, the credible decision score of the calibration sample is calculated based on that weight combination, and the acceptance threshold corresponding to that weight combination is simultaneously determined using a category-budget two-dimensional threshold matrix learning method. Among the candidate weight combinations that meet the preset upper limit of the error acceptance rate, the weight combination with the fewest average decision rounds is selected as the final credible score weight.

[0023] As a further improvement of the present invention, the acceptance threshold adopts a category-budget two-dimensional threshold matrix. It means that, among them, Indicates candidate category The acceptance threshold corresponding to the t-th observation budget; the category-budget two-dimensional threshold matrix The data was obtained through offline learning based on a multi-network environment calibration dataset and compared with the environment calibration item. These are used together to determine the acceptance of the current candidate category under the current observation budget, incorporating prior differences under different network environments, link conditions, or deployment scenarios into the reliable determination process. During the offline calibration learning phase, for each candidate category... and observation budget The process involves calculating the credible decision score for each calibration sample and iterating through the candidate acceptance thresholds. For a given candidate acceptance threshold, samples with a credible decision score greater than or equal to that threshold are considered accepted samples. The proportion of accepted samples whose candidate class differs from the true class to the total number of accepted samples is defined as the false acceptance rate. The false acceptance rate is set to a preset upper limit. From the candidate thresholds, the threshold that can accept a larger number of calibration samples is selected as the threshold. The above process is repeated for each candidate category and each observation budget to obtain a two-dimensional threshold matrix of category-budget. The preset error acceptance rate upper limit. Pre-set the security requirements according to the actual deployment scenario; for deployment scenarios with high security requirements, a smaller upper limit for the error acceptance rate can be set to obtain stricter early acceptance conditions.

[0024] As a further improvement of the present invention, when the credible decision score is less than the acceptance threshold and the maximum observation budget has not yet been reached, the current observation budget is adjusted from... Updated to the next adjacent observation budget in the preset observation budget sequence. In this context, each observation budget represents the upper limit of observations accumulated from the starting position of the self-flowing analysis object. The newly added observations in this round are... To characterize the degree of divergence among the view sketches in the current round, the class probability distribution output by the i-th view sketch under the t-th observation budget is used. Probability distribution of the current fusion category The Jensen-Shannon divergence between the two elements determines the degree of divergence in the view sketch. :

[0025] ;

[0026] in, Let be the Jensen-Shannon divergence function; 1 ≤ i ≤ K, where K is the observation budget. The number of view sketches is equal to the number of view constructors;

[0027] The degree of divergence of the sketches in each view The average is used to obtain the overall degree of disagreement in the current round. :

[0028] ;

[0029] make ,in, The index of the view sketch with the largest divergence under the t-th observation budget, and the index of the t-th observation budget. The view sketch is determined as the main divergence view sketch for the current round; the main divergence view sketch is used to identify the feature dimension with the most significant divergence in the current classification result.

[0030] Finally, after each round of additional observations is completed, the steps of prefix observation, view sketch generation, classification processing, and confidence assessment are repeated until the acceptance threshold is met or the maximum observation budget is reached.

[0031] As a further improvement to the present invention, the specific steps of the upgrade process are as follows:

[0032] First, when the trusted decision score of the encrypted stream to be classified is still less than the acceptance threshold under the maximum observation budget, the encrypted stream to be classified is determined to be an object that the edge node cannot reliably classify.

[0033] Then, based on the probability distribution of fusion categories under the current observation budget, the view Figure 1 The system evaluates consistency, stability between adjacent observation budgets, and a reliable decision score, and outputs at least one of the following: low-reliability result, unknown result, or risk result, as a preliminary judgment conclusion on the edge side; wherein, when the current candidate category belongs to a pre-configured risk category or a local security policy is triggered, the preliminary judgment conclusion can be marked as a risk result.

[0034] Subsequently, the multi-view sketches generated during the current observation process are compressed and encoded, and the compressed and encoded view sketches, along with the decision trajectory and session metadata, are combined into an upgrade processing package; wherein, the decision trajectory includes the top q candidate categories and their corresponding probabilities corresponding to the probability distribution of each round of fusion categories, the credible decision scores of each round, and the observation budget and timestamps corresponding to each round;

[0035] Where q is a positive integer. In each round of observation, the category probability distributions corresponding to the sketches of each view in that round are first fused to obtain the fused category probability distribution of that round. Then, each candidate category in the fused category probability distribution is sorted from largest to smallest according to its probability value. The top q candidate categories are selected, and the probability values ​​of these q candidate categories in the fused category probability distribution of that round are output.

[0036] The upgrade processing package may also include a hash digest, which is generated by compression encoding based on the corresponding view sketch and is used to verify the integrity of the uploaded data.

[0037] Next, the upgraded processing package is sent to the upper-level analysis node, which then uses one or more of the following methods to further determine the encrypted stream to be classified: invoking a high-complexity classification model for further review, rule engine analysis, threat intelligence analysis, historical behavior comparison, or manual review. The high-complexity classification model can be a Transformer classification model, a deep residual network (ResNet), a graph convolutional network (GCN), or a graph attention network (GAT), etc. Preferably, the upper-level analysis node can also combine global traffic context, cross-session association information, terminal profile information, or security policy information to jointly analyze low-trust traffic uploaded by edge nodes, thereby improving the identification accuracy in complex scenarios.

[0038] Finally, based on the final judgment result returned by the superior analysis node, corresponding actions are executed, including but not limited to: result feedback, traffic marking, alarm reporting, image retention, access control tightening, zero-trust secondary verification or policy approval, thereby completing the closed-loop upgrade process from the edge side to the center side.

[0039] A multi-view Figure 1 A cryptographic stream classification system with consistency and progressive observation control includes:

[0040] The session aggregation module is configured to aggregate the collected encrypted streams to be classified, forming stream-level analysis objects;

[0041] The budget setting module is configured to set an observation budget sequence, which consists of multiple incremental observation budgets, and select the first observation budget;

[0042] The prefix observation module is configured to extract the prefix message subsequence of the flow-level analysis object under the current observation budget;

[0043] The view sketch generation module is configured to generate at least two view sketches based on the prefix message subsequence;

[0044] The classification module is configured to input the view sketches into the corresponding classification models to obtain the category probability distribution and embedding representation of each view sketch; obtain the fusion category probability distribution under the current observation budget based on the category probability distribution of each view sketch, and determine the candidate category with the highest probability in the fusion category probability distribution as the current candidate category;

[0045] The credibility assessment module is configured to calculate a credibility decision score under the current observation budget based on the category probability distribution and embedding representation, wherein the credibility decision score is composed of the candidate category confidence margin, the viewpoint, and the confidence margin. Figure 1 Consistency, stability between adjacent observation budgets, environmental calibration terms, and instability penalty terms are jointly used to determine the candidate category; among them, the candidate category confidence margin, apparent value, and other factors are considered. Figure 1 Consistency, stability between adjacent observation budgets, and instability penalty terms are calculated based on current and historical classification outputs. The environmental calibration terms and the corresponding weights of each evaluation term are determined offline based on a multi-network environmental calibration dataset.

[0046] The control module is configured to perform control based on a comparison between a trusted decision score and an acceptance threshold corresponding to the current candidate category under the current observation budget: when the trusted decision score is greater than or equal to the acceptance threshold, the control outputs the current candidate category as the classification result and stops further observation; when the trusted decision score is less than the acceptance threshold and the maximum observation budget has not been reached, the current observation budget is updated to the next adjacent observation budget in a preset observation budget sequence, and the control extends the prefix message subsequence; when the trusted decision score is still less than the acceptance threshold when the maximum observation budget is reached, the control generates an upgrade disposal package and sends it to the upper-level analysis node; and after receiving the final judgment result returned by the upper-level analysis node, the control generates a corresponding disposal instruction based on the final judgment result and calls the policy execution interface to execute or issue the disposal instruction.

[0047] As a further improvement of the present invention, the observation budget is characterized by at least one of the following: the number of prefix messages, the length of the prefix time window, the number of message direction changes, and the number of burst units. The budget setting module sets an incremental observation budget sequence for the adopted budget representation method, and selects or adjusts the observation budget sequence within the range of candidate observation budget values ​​that have been pre-calibrated offline, in combination with the network scenario, link status, service type, and device computing power conditions, so that each adjusted observation budget has a corresponding acceptance threshold configuration.

[0048] As a further improvement of the present invention, the classification module extracts features using a lightweight convolutional network, a lightweight sequence model, or a graph encoder, based on the feature dimensions of different view sketches, as well as the deployment scenario and computing power conditions, and outputs the class probability distribution and embedding representation of each view sketch; wherein:

[0049] Lightweight convolutional networks are used to extract local pattern features from view sketches;

[0050] Lightweight sequence models are used to model temporal dependencies in prefix traffic;

[0051] Graph encoders are used to learn state transition relationships and structural association features;

[0052] Alternatively, each view sketch can share the underlying encoder parameters, or some view sketches can share the underlying encoder parameters and set independent classification heads respectively, so as to retain the independent class probability distribution and embedding representation of each view sketch while reducing the number of model parameters and computational overhead.

[0053] The category probability distribution and intermediate embedding representation corresponding to each output view sketch are used as input for subsequent calculation of the credible decision score.

[0054] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0055] (1) This invention does not force a hard judgment on encrypted streams with insufficient evidence, but instead uses multiple views. Figure 1 The comprehensive assessment of the reliability of consistency and budget stability can significantly reduce the risk of misclassification in scenarios with insufficient evidence; it supports the output of classification results in advance when the reliability conditions are met, shortening the average decision latency and reducing edge computing power consumption without sacrificing reliability; through a three-level mechanism of "continued observation - re-judgment - escalation of handling when necessary", it improves the system's adaptability to environmental changes, unknown traffic and complex mixed services.

[0056] (2) This invention does not require uploading the original traffic content during upgrade processing. Instead, it uploads upgrade processing information including view sketch compression encoding and decision trajectory, which helps reduce the data transmission overhead between edge nodes and upper-level analysis nodes and facilitates deployment in edge nodes, gateway devices, zero-trust access devices or traffic auditing devices.

[0057] (3) This invention incorporates feature characterization, reliable assessment, observation and control, and upgrade disposal into a unified closed loop, which has strong engineering applicability and scalability. Attached Figure Description

[0058] Figure 1 This is a flowchart of an embodiment of the present invention.

[0059] Figure 2 This is a schematic diagram of multi-view sketch generation in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the credible decision score and progressive observation control in an embodiment of the present invention;

[0061] Figure 4 This is a diagram illustrating the collaborative upgrade and processing architecture between edge nodes and upper-level analysis nodes in an embodiment of the present invention.

[0062] Figure 5 This is a schematic diagram illustrating the separate deployment of the training phase and the online inference phase in an embodiment of the present invention;

[0063] Figure 6 This is a system block diagram in an embodiment of the present invention. Detailed Implementation

[0064] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0065] This invention provides a multi-view Figure 1 For a cryptographic stream classification method based on consistency and asymptotic observation control, see [link to relevant documentation]. Figure 1 ,include:

[0066] Step S1: Collect encrypted flows to be classified at edge nodes, and aggregate the encrypted flows to be classified based on 5-tuples, connection identifiers, or session keys. Integrate network packets belonging to the same communication connection or the same session into flow-level analysis objects. Configure an incremental observation budget sequence for each flow-level analysis object, consisting of a first budget sequence with an increasing prefix packet count (e.g., prefix packet count increasing to 4, 8, 16, 32, 64) and / or a second budget sequence with an increasing prefix time window length (e.g., prefix time window length increasing to 20ms, 50ms, 100ms, 200ms, 500ms). 'n' represents the observation budget quantity, and the current observation budget sequence can be selected or adjusted within the range of candidate observation budget values ​​that have undergone pre-comparison and offline calibration, taking into account network scenarios, link states, service types, and device computing power. For example, when using the number of prefix packets as the observation budget, it can be set... At this time... This indicates that the first round of observation uses the first four messages of the encrypted stream from the starting position; if the first round does not meet the acceptance conditions and enters... In the second round, the first eight packets from the start of the flow are used. Therefore, the prefix packet subsequences for each round can be determined. ,1≤m≤n-1.

[0067] Step S2, let ,Will As the current observation budget, a prefix message subsequence is extracted from the starting position of the flow level analysis object according to the current observation budget. In the subsequent t-th round of observations, based on the current observation budget... Extract the corresponding prefix message subsequence from the starting position of the analysis object at the same flow level. ,in Each round determines only one current observation budget and generates a prefix message subsequence corresponding to that observation budget. All view sketches in the same round are generated based on this prefix message subsequence.

[0068] Step S3: For the prefix message subsequence Preprocessing is performed, including one or more of the following: direction marking, time alignment, anomaly filtering, or length discretization; for the preprocessed... K types of view constructors are used respectively. Generate a collection of multi-view sketches ,in This represents the i-th view sketch obtained under the t-th observation budget, where K ≥ 2; 1 ≤ i ≤ K. In a preferred embodiment, as shown... Figure 2 As shown, the prefix message subsequence under the t-th observation budget is... As a common input for various view construction methods, corresponding view sketches are generated from different feature dimensions, where the view sketch types include:

[0069] The length-direction sketch is used to map the length of each message to a preset length range and combine it with the uplink and downlink directions to form a length-direction sequence arranged in the order of messages, representing the variation characteristics of traffic in the length and direction dimensions in the form of sequence stripes;

[0070] A burst sketch is used to merge adjacent messages that are consecutive in time and in the same direction into a burst unit, and to record the number of messages, the cumulative number of bytes, and the duration of each burst unit in the order of their appearance, forming a burst unit sequence;

[0071] The time normalization sketch is used to normalize the arrival time interval between adjacent packets and map the normalized time information to a unified time scale to form a normalized time series, so as to reduce the impact of link jitter and latency changes.

[0072] A state transition sketch is used to encode the state of each message according to the message direction state, length state, and time state, and to count the transition relationship between adjacent states to form a state transition matrix to represent the dynamic change characteristics of the encrypted stream between different states.

[0073] The packet loss disturbance sketch is used to perform controlled masking or sampling on the prefix message subsequence according to preset rules, and to form a robust auxiliary view after disturbance based on the retained message information, so as to simulate input changes under conditions of slight packet loss, sampling error or link noise.

[0074] Step S4: Each view sketch is fed into a view classifier composed of a lightweight convolutional network, a lightweight sequence model, or a graph encoder, according to its feature dimensions, to obtain the class probability distribution and embedding representation of each view sketch. To facilitate consistent calculation of cross-view embedding representations, the embedding representations output by each view classification model are uniformly set to the same preset dimension and normalized before calculating cosine similarity. In a preferred embodiment, each view shares the underlying encoder parameters, using an independent classification head only at the output end to reduce the computational overhead of edge devices.

[0075] Let the class probability distribution of the i-th view sketch under the t-th observation budget be... In a preferred embodiment, the class probability distributions of the K view sketches are averaged and fused to obtain the fused class probability distribution under the current observation budget:

[0076] ;

[0077] The candidate category with the highest probability in the fusion category probability distribution is determined as the current candidate category:

[0078] ;

[0079] Then, the candidate categories in the fusion category probability distribution are sorted from largest to smallest probability to obtain the top q candidate categories and their corresponding probabilities under the current observation budget.

[0080] Step S5: Class probability distribution based on each view sketch and embedding representation Calculate the current observation budget Credible decision score Preferably, based on the fusion category probability distribution obtained in step S4 and the current candidate category... Calculate the following credibility evaluation items respectively:

[0081] Candidate category confidence margin Determined based on the difference between the probability of the first category and the probability of the second category in the current fusion category probability distribution;

[0082] See Figure 1 Desire The probability distribution of each view category is determined based on the difference between the probability distribution of each view category and the probability distribution of the fused category, as well as the similarity between the embedding representations of each view. The closer the probability distributions of each view are and the more similar their embedding representations are, the better. The larger;

[0083] Stability between adjacent observation budgets Based on the probability distribution of fusion categories under the current observation budget Fusion category probability distribution under the previous observation budget The difference between the two is determined; the smaller the difference, the better. The larger. When At that time, because there was no previous observation budget, The acceptance criteria for the first observation budget are individually calibrated offline using the corresponding thresholds.

[0084] Environmental calibration items Based on the current network environment, current candidate category, and current observation budget, environmental reliability parameters are queried from the offline multi-network environment calibration dataset and normalized to a preset numerical range. Specifically, the multi-network environment calibration dataset is grouped according to network environment, candidate category, and observation budget. For network environment... Candidate categories and observation budget The environmental reliability parameters are obtained by calculating the proportion of samples whose current candidate category matches the true category among the corresponding calibration samples. During online inference, the current network environment must be considered. Current candidate category and current observation budget Query the corresponding current environment reliability parameters and use it as the current environmental calibration item. .

[0085] Instability penalty item The candidate category is determined based on whether it has changed relative to the candidate category in the previous observation budget and the category changes in recent observation rounds; the more frequently the candidate category changes, the more likely it is to change. The larger; when season ;

[0086] In a preferred embodiment, the differences between view category probability distributions and between adjacent observation budget fusion category probability distributions are measured using Jensen-Shannon divergence, and the similarity between view embedding representations is measured using cosine similarity, with the results normalized to [value missing]. The interval; for the instability penalty term, it is set to 0 when the current candidate category is consistent with the candidate category under the previous observation budget, and 1 when they are inconsistent; for the first observation, since there is no previous observation budget, let , For vision Figure 1 Desire The probability consistency value is obtained by subtracting the average Jensen-Shannon divergence between the probability distributions of each view category and the fused category from 1; the embedding consistency value is obtained by calculating the average cosine similarity between each pair of view embedding representations; and the probability consistency value and the embedding consistency value are then arithmetically averaged to obtain the view consistency value. Figure 1 Desire .

[0087] According to the following formula:

[0088] ;

[0089] The current credible decision score is calculated. Wherein, Let be the credible decision score under the budget of the t-th observation; This represents the confidence margin of the candidate category under the budget of the t-th observation. Represents the view under the budget of the t-th observation. Figure 1 Devotion, This indicates the stability between adjacent observation budgets under the t-th observation budget. This represents the environmental calibration term under the t-th observation budget. Let represent the instability penalty term under the budget of the t-th observation. Where α, β, γ, η, and δ are respectively... , , , , The corresponding non-negative weights are determined by parameter search or constraint optimization based on a multi-network environment calibration dataset according to the aforementioned offline calibration method, and are deployed as trusted scoring parameters to online inference nodes.

[0090] Step S6: Based on the current candidate category and current observation budget From the category-budget two-dimensional threshold matrix Obtain the corresponding acceptance threshold from Due to environmental calibration items It has been included in the credible decision score calculation formula. The calculations are therefore included in the acceptance criteria based on the calibration results corresponding to the current network environment. At that time, the current candidate category will be... The classification result of the stream is output, and further observation is stopped to achieve early stopping classification.

[0091] like and If the current flow-level analysis object and its session aggregation state remain unchanged, the current observation budget will be changed from... Update to the budget of the next adjacent observation When the flow-level analysis object cumulatively meets the following conditions: Under the corresponding observation conditions, the expanded prefix message subsequence is extracted from the starting position of the analysis object at the same flow level. And repeat steps S2 to S6 without re-executing the session aggregation in step S1.

[0092] Step S7, if and If the output is low-confidence, unknown, or risky, an upgrade handling package will be generated. The upgrade processing package It should include at least: compressed encoding of the view sketch, the top q candidate categories in each round and their corresponding probabilities, the confidence score trajectory, the budget trajectory, and session metadata; it may also include a hash digest, which will then be... The data is sent to the higher-level analysis node, which then performs one or more further judgments, including calling a high-complexity classification model for further review, rule engine analysis, threat intelligence analysis, and manual verification. A high-complexity classification model, in contrast to the lightweight classification model deployed on edge nodes, refers to a classification model with relatively high computational complexity and model size, used for further reviewing low-confidence classification results at the edge. This high-complexity classification model can employ Transformer classification models, deep residual networks (ResNet), graph convolutional networks (GCN), or graph attention networks (GAT), etc.

[0093] Step S8: Execute corresponding actions based on the final judgment result, including but not limited to traffic marking, alarms, image retention, fine-grained access control, zero-trust secondary verification, or policy-based approval.

[0094] This invention introduces multi-view into the encrypted traffic classification process. Figure 1 Consistency assessment, budget stability assessment, and progressive observation budget control mechanism have enabled a closed-loop classification process that allows for "rapid classification when early cessation is possible, continued observation when evidence is insufficient, and escalation of handling when still unreliable." This process can improve the credibility, robustness, and deployment efficiency of the classification system without the need for decryption.

[0095] like Figure 3 As shown, in the credible decision score and progressive observation control process (steps S5-S8), the class probability distributions of the multiple views (View1-ViewK) output by the split-view classifier and the embedding representations of the multiple views (V1-VK) are used as inputs for the credible assessment. The credible assessment obtains the candidate class confidence margins under the current observation budget. ,See Figure 1 Desire Stability between adjacent observation budgets Environmental calibration items and instability penalty items The current credible decision score is obtained according to the aforementioned credible decision score calculation method. Subsequently, based on the current candidate category and current observation budget Obtain the corresponding acceptance threshold from the category-budget two-dimensional threshold matrix. And compare credible decision scores With the acceptance threshold .when When the current candidate category is entered, the classification result is accepted and the classification result is output; when If the maximum observation budget has not yet been reached, proceed to path two and update the current observation budget to the next adjacent observation budget. And proceed to the next round of prefix observations and credibility determination; when the maximum observation budget is reached... If the result is still less than the corresponding acceptance threshold, proceed to path three, mark the traffic as a low-confidence result, unknown result, or risky result, and enter the upgrade handling process.

[0096] like Figure 4 As shown, this embodiment employs a collaborative deployment approach between edge nodes and upper-level analysis nodes. Edge devices, gateways, or audit nodes are responsible for performing traffic collection, multi-view sketch generation, lightweight classification, and trust assessment. For traffic that meets the acceptance criteria, classification results are directly generated at the edge. For traffic that does not meet the acceptance criteria even under the maximum observation budget, the edge node generates an escalation handling package. This package includes compressed encoding of the view sketch and optional hash digests, the top q candidate categories in each round and their corresponding probabilities, trust decision scores, budget trajectories, and session metadata, and sends it to the upper-level analysis node or security platform. The upper-level analysis node further determines the outcome based on the received escalation handling package by employing one or more of the following methods: invoking a high-complexity classification model for re-judgment, using a rule engine, threat intelligence analysis, and manual review.

[0097] After obtaining the final judgment result, the final judgment result is returned to the edge node. The control module of the edge node generates a corresponding handling instruction based on the final judgment result, and executes the instruction by calling the policy execution interface or sends the handling instruction to the corresponding network device.

[0098] like Figure 5As shown, this embodiment separates the offline training phase from the online inference phase. In the offline training phase, the classification model corresponding to each view is trained using the training dataset to obtain the classification model parameters for multiple views. The classification model is then compressed and exported as a model file based on edge deployment conditions. Simultaneously, the trained classification model is applied to a multi-network environment calibration dataset for calibration set inference, obtaining the classification output under each network environment, candidate category, and observation budget. Based on this, the confidence scoring parameters and the category-budget two-dimensional threshold matrix are determined according to the aforementioned offline calibration method, forming the corresponding confidence scoring parameters and threshold configuration file. In the online inference phase, the edge nodes load the model file and the confidence scoring parameters and threshold configuration file, and sequentially execute edge acquisition, prefix observation, multi-view inference, and confidence scoring. When the corresponding acceptance threshold is met, early stopping is executed; when the acceptance condition is not met and the maximum observation budget has not yet been reached, the next round of observation continues; when the maximum observation budget is reached but the acceptance condition is still not met, an upgrade process is executed. Therefore, the online inference phase does not require retraining the classification model or learning the threshold online.

[0099] In another alternative embodiment, edge nodes are deployed in enterprise network egress gateways, software-defined wide area network (SD-WAN) nodes, industrial internet security isolation devices, campus network egress gateways, or data center bypass auditing devices; upper-level analysis nodes are deployed in local security analysis platforms or cloud-based security operations centers. For business traffic that can obtain highly reliable decisions within a short observation budget, this invention can output classification results in advance, reducing overall detection latency; for traffic that cannot be reliably identified within a short prefix, this invention avoids misjudgment through continued observation and upgraded handling mechanisms, thereby improving its practicality in real network environments.

[0100] This invention also provides a multi-view... Figure 1 A cryptographic stream classification system for consistency and asymptotic observation control, see [link to documentation]. Figure 6 ,include:

[0101] The session aggregation module is configured to aggregate the collected encrypted streams to be classified, forming stream-level analysis objects;

[0102] The budget setting module is configured to set an observation budget sequence, which consists of multiple incremental observation budgets, and select the first observation budget;

[0103] The prefix observation module is configured to extract the prefix message subsequence of the flow-level analysis object under the current observation budget;

[0104] The view sketch generation module is configured to generate at least two view sketches based on the prefix message subsequence;

[0105] The classification module is configured to input the view sketches into the corresponding classification models to obtain the category probability distribution and embedding representation of each view sketch; obtain the fusion category probability distribution under the current observation budget based on the category probability distribution of each view sketch, and determine the candidate category with the highest probability in the fusion category probability distribution as the current candidate category;

[0106] The credibility assessment module is configured to calculate a credibility decision score under the current observation budget based on the category probability distribution and embedding representation, wherein the credibility decision score is composed of the candidate category confidence margin, the viewpoint, and the confidence margin. Figure 1 Consistency, stability between adjacent observation budgets, environmental calibration terms, and instability penalty terms are jointly used to determine the candidate category; among them, the candidate category confidence margin, apparent value, and other factors are considered. Figure 1 Consistency, stability between adjacent observation budgets, and instability penalty terms are calculated based on current and historical classification outputs. The environmental calibration terms and the corresponding weights of each evaluation term are determined offline based on a multi-network environmental calibration dataset.

[0107] The control module is configured to perform control based on a comparison between a trusted decision score and an acceptance threshold corresponding to the current candidate category under the current observation budget: when the trusted decision score is greater than or equal to the acceptance threshold, the control outputs the current candidate category as the classification result and stops further observation; when the trusted decision score is less than the acceptance threshold and the maximum observation budget has not been reached, the current observation budget is updated to the next adjacent observation budget in a preset observation budget sequence, and the control extends the prefix message subsequence; when the trusted decision score is still less than the acceptance threshold when the maximum observation budget is reached, the control generates an upgrade disposal package and sends it to the upper-level analysis node; and after receiving the final judgment result returned by the upper-level analysis node, the control generates a corresponding disposal instruction based on the final judgment result and calls the policy execution interface to execute or issue the disposal instruction.

[0108] Although the present invention has been described herein with reference to illustrative embodiments, the above embodiments are merely preferred embodiments of the present invention, and the implementation of the present invention is not limited to the above embodiments. It should be understood that those skilled in the art can devise many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A method of encrypted stream classification for multi-view consistency and progressive observation control, characterized in that, include: Step S1: Aggregate the collected encrypted streams to be classified to form stream-level analysis objects; Set an observation budget sequence, which consists of multiple incremental observation budgets, and select the first observation budget to proceed to the next step; Step S2: Extract the prefix message subsequence of the flow level analysis object under the current observation budget; Step S3: Generate at least two view sketches based on the prefix message subsequence; Step S4: Input the view sketches into the corresponding classification models to obtain the category probability distribution and embedding representation of each view sketch; obtain the fusion category probability distribution under the current observation budget based on the category probability distribution of each view sketch, and determine the candidate category with the highest probability in the fusion category probability distribution as the current candidate category; Step S5: Calculate the credible decision score under the current observation budget based on the category probability distribution and embedding representation. The credible decision score is jointly determined by the candidate category confidence margin, view consistency, stability between adjacent observation budgets, environmental calibration term, and instability penalty term. Step S6, Judgment: When the credible decision score is greater than or equal to the acceptance threshold corresponding to the current candidate category under the current observation budget, the classification result is output and observation is stopped. When the credible decision score is less than the acceptance threshold and the maximum observation budget has not been reached, select the next observation budget adjacent to the current observation budget in the observation budget sequence and return to step S2; If the confidence decision score is still less than the acceptance threshold when the maximum observation budget is reached, a low confidence result, an unknown result, or a risk result is output, and an upgrade handling package is sent to the superior analysis node for upgrade handling. The system also receives the final judgment result returned by the superior analysis node and executes the corresponding handling action.

2. The encrypted flow classification method for multi-view consistency and progressive observation control according to claim 1, characterized in that, The observation budget is characterized by at least one of the following: the number of prefix messages, the prefix time window length, the number of message direction changes, and the number of burst units; an observation budget sequence is set according to the budget representation method adopted. : ,in, Let t be the budget for the t-th observation, where t is the number of the budget for the observation, 1≤t≤n, and n is the number of budgets for the observation.

3. The encrypted flow classification method for multi-view consistency and progressive observation control according to claim 1, characterized in that, The view sketch includes at least two of the following: length-direction sketch, burst sketch, time-normalized sketch, state transition sketch, and anti-packet loss disturbance sketch.

4. The encrypted flow classification method for multi-view consistency and progressive observation control according to claim 1, characterized in that, The credible decision score is determined by the following formula: ; in, The score represents the credible decision score. Confidence margin for candidate categories, For view consistency, For the stability between adjacent observation budgets, For environmental calibration items, As an unstable penalty item, Preset weights for the confidence margin of candidate categories; Preset weights for view consistency; Preset weights for stability between adjacent observation budgets; Preset weights for environmental calibration items; The preset weights for the unstable penalty terms.

5. The encrypted flow classification method for multi-view consistency and progressive observation control according to claim 4, characterized in that, The acceptance threshold is adopted using a category-budget two-dimensional threshold matrix. express: ; in, Indicates candidate category The acceptance threshold corresponding to the t-th observation budget, the category-budget two-dimensional threshold matrix The data was obtained through offline learning based on a multi-network environment calibration dataset and compared with the environment calibration item. Together they are used to determine the acceptance of the current candidate category under the current observation budget.

6. The encrypted flow classification method for multi-view consistency and progressive observation control according to claim 1, characterized in that, The upgrade processing package includes compressed encoding of the view sketch, decision trajectory, and session metadata. The decision trajectory includes candidate categories corresponding to the top q probability distributions of each round's fusion category probability distribution, the probability value of the candidate category in the corresponding fusion category probability distribution, the credible decision score for each round, the observation budget and timestamp for each round, where q is a positive integer.

7. The encrypted flow classification method for multi-view consistency and progressive observation control according to claim 1, characterized in that, The upgrade process includes one or more of the following: calling a high-complexity classification model for further review, rule engine analysis, threat intelligence analysis, historical behavior comparison, and manual review; the high-complexity classification model is a Transformer classification model, a deep residual network ResNet, a graph convolutional network GCN, or a graph attention network GAT.

8. A cryptographic stream classification system with multi-view consistency and progressive observation control, characterized in that, include: The session aggregation module is configured to aggregate the collected encrypted streams to be classified, forming stream-level analysis objects; The budget setting module is configured to set an observation budget sequence, which consists of multiple incremental observation budgets, and select the first observation budget; The prefix observation module is configured to extract the prefix message subsequence of the flow-level analysis object under the current observation budget; The view sketch generation module is configured to generate at least two view sketches based on the prefix message subsequence; The classification module is configured to input the view sketches into the corresponding classification models to obtain the category probability distribution and embedding representation of each view sketch; obtain the fusion category probability distribution under the current observation budget based on the category probability distribution of each view sketch, and determine the candidate category with the highest probability in the fusion category probability distribution as the current candidate category; The credibility assessment module is configured to calculate a credibility decision score under the current observation budget based on the category probability distribution and the embedding representation. The credibility decision score is jointly determined by the candidate category confidence margin, view consistency, stability between adjacent observation budgets, environmental calibration term, and instability penalty term. The control module is configured to perform control based on a comparison between a trusted decision score and an acceptance threshold corresponding to the current candidate category under the current observation budget: when the trusted decision score is greater than or equal to the acceptance threshold, the current candidate category is output as the classification result and observation is stopped; when the trusted decision score is less than the acceptance threshold and the maximum observation budget has not been reached, the current observation budget is updated to the next adjacent observation budget in the preset observation budget sequence, and the prefix observation module is controlled to expand the prefix message subsequence. When the credible decision score is still less than the acceptance threshold when the maximum observation budget is reached, an upgrade disposal package is generated and sent to the superior analysis node. Upon receiving the final judgment result returned by the superior analysis node, the system generates a corresponding handling instruction based on the final judgment result and calls the strategy execution interface to execute or issue the handling instruction.

9. The encrypted stream classification system with multi-view consistency and progressive observation control according to claim 8, characterized in that, The observation budget is characterized by at least one of the following: the number of prefix messages, the length of the prefix time window, the number of message direction changes, and the number of burst units. The budget setting module sets an incremental observation budget sequence for the adopted budget representation method, and selects or adjusts the observation budget sequence within the range of candidate observation budget values ​​that have been pre-calibrated offline, in combination with the network scenario, link status, service type, and device computing power conditions, so that each adjusted observation budget has a corresponding acceptance threshold configuration.

10. The encrypted stream classification system with multi-view consistency and progressive observation control according to claim 8, characterized in that, The classification module extracts features using lightweight convolutional networks, lightweight sequence models, or graph encoders based on the feature dimensions of different view sketches, as well as the deployment scenario and computing power conditions, and outputs the class probability distribution and embedding representation of each view sketch; or the underlying encoder parameters are shared or partially shared by each view sketch, and independent classification heads are set for each view sketch, outputting the class probability distribution and embedding representation of each view sketch.

Citation Information

Patent Citations

  • Encrypted traffic classification method based on multi-view heterogeneous graph model

    CN119719895A

  • Network intrusion detection method for enhancing minority class data identification capability

    CN120415855A