Encrypted malicious traffic detection system based on deep learning
By constructing a multi-module collaborative deep learning detection system, the problems of scarce labeled data, susceptibility to feature interference, and vulnerability to adversarial attacks in encrypted malicious traffic detection are solved, achieving efficient, accurate, and robust malicious traffic detection in dynamic network environments.
Patent Information
- Application Number
- CN202511585042.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-31
- Publication Date
- 2026-02-10
AI Technical Summary
Existing deep learning-based methods for detecting encrypted malicious traffic face challenges such as scarce labeled data, low feature effectiveness, susceptibility to adversarial attacks, and poor environmental adaptability, especially in dynamic network environments where it is difficult to effectively detect encrypted traffic.
A deep learning-based encrypted malicious traffic detection system is constructed, including an intelligent data acquisition module, a feature extraction and filtering module, an intelligent data analysis module, an annotation quality verification module, a metadata enhancement detection module, an adversarial disturbance detection module, a dynamic collaborative control module, and a real-time online feedback module. Through the collaboration of multiple modules, automated data annotation, robust feature extraction, adversarial attack resistance, and dynamic environment adaptation are achieved.
It improves the accuracy and robustness of encrypted traffic detection, effectively identifies malicious traffic in dynamic network environments, reduces the impact of adversarial attacks, and ensures the stability and adaptability of the system in different encryption scenarios.
Smart Images
Figure CN121508927A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to the technical field of encrypted network flow security detection, in particular to an encrypted malicious flow detection system based on deep learning. BACKGROUND
[0002] With the wide popularity of encryption protocols such as TLS, network flow encryption has become the norm, which, although guaranteeing user privacy, has brought severe challenges to malicious flow detection. Traditional deep learning-based detection methods face many bottlenecks in practical application: first, high-quality labeled data is difficult to obtain, especially in scenarios where new attacks occur frequently, manual labeling is costly and seriously lags behind threat evolution; although semi-supervised learning can alleviate the dependence on labeling, the accuracy of the pseudo-labels it relies on lacks a reliable verification mechanism. Second, encryption technology hides the plaintext content, forcing detection to rely on metadata (such as packet length, timing, etc.), which is easily disturbed by noise and has weak cross-protocol migration ability. In addition, deep learning models are vulnerable to adversarial attacks, and attackers can significantly reduce the detection rate by applying small perturbations, while existing defense measures lack universality. Finally, the real network environment is complex and variable, and the encryption stack is diverse (such as different TLS versions and Internet of Things devices), and the adaptation ability and robustness of existing models in dynamic environments have not been fully verified. Therefore, there is an urgent need for a new detection scheme that can systematically solve the above problems. SUMMARY
[0003] (I) Technical problems solved In view of the deficiencies of the prior art, the present application provides an encrypted malicious flow detection system based on deep learning, which has the advantages of automatic data labeling verification, robust feature extraction, anti-adversarial attack, and dynamic environment adaptation, and solves the problems of labeled data scarcity, low feature effectiveness, model vulnerability to adversarial attack, and poor environmental adaptability in encrypted flow detection.
[0004] (II) Technical solutions To achieve the above-mentioned purpose, the present application provides the following technical solutions: an encrypted malicious flow detection system based on deep learning, comprising an intelligent data acquisition module, a feature extraction and filtering module, an intelligent data analysis module, a labeling quality verification module, a metadata enhancement detection module, an adversarial perturbation detection module, a dynamic cooperative control module, a real-time online feedback module, and a full-scene effectiveness verification module. The intelligent data acquisition module is responsible for collecting raw metadata. The feature extraction and filtering module constructs high-dimensional features from raw metadata and filters noise to extract robust feature representations across protocols and environments. The intelligent data analysis module receives feature input from the feature extraction and filtering module and calculates the labeling data quality score Average discrimination of multi-dimensional features Feature perturbation threshold for adversarial attack determination ; The annotation quality verification module is based on annotation data quality scoring. The calculation formula is used to compare pseudo-labels with real labels and to verify the consistency of labels across iteration rounds. The feature validity determination module is based on the average discrimination of multi-dimensional features. The calculation formula is used to compare cross-protocol feature distinguishability and verify feature-tag mutual information. The adversarial disturbance detection module is based on the characteristic disturbance threshold for adversarial attack determination. The calculation formula is used to calculate the feature perturbation of the current traffic sample and compare it with the threshold. The dynamic collaborative control module generates control commands based on the output results of the annotation quality verification module, feature validity determination module, and anti-disturbance detection module, and issues them to the intelligent data acquisition module and feature extraction and filtering module to achieve adaptive system parameters. The real-time online feedback module monitors the operating status and output indicators of each module online through a real-time data stream monitoring and indicator feedback mechanism, and performs real-time capture of abnormal information and online updating of strategies. The full-scenario validity verification module performs full-process testing on the system's annotation quality, feature validity, adversarial defense capabilities, and dynamic adaptation capabilities under multiple encryption scenarios, verifying whether all issues have been resolved.
[0005] Preferably, the intelligent data acquisition module includes a metadata acquisition unit, a semi-supervised annotation unit, and an adversarial enhancement unit; The metadata collection unit extracts metadata such as packet length, timing, direction, and plaintext information of the TLS handshake from the traffic. The semi-supervised labeling unit generates pseudo-labels for unlabeled data based on a pre-trained teacher model. The adversarial enhancement unit applies controllable adversarial perturbations to existing labeled data to generate adversarial samples for training.
[0006] Preferably, the feature extraction and filtering module includes a spatiotemporal feature construction unit and a noise filtering unit; The spatiotemporal feature construction unit converts the packet-length sequence and interaction interval into feature vectors; The noise filtering unit filters low-value or high-variance features based on feature importance scores.
[0007] Preferably, the intelligent data analysis module includes a labeled data quality assessment unit, a feature discrimination calculation unit, and an adversarial disturbance threshold determination unit.
[0008] Preferably, the annotation data quality assessment unit calculates the annotation data quality score. The calculation formula is as follows: ; In the formula, This indicates the quality score of the labeled data. The weights representing the accuracy of pseudo-labels The weight representing the consistency of pseudo-labels across rounds. This represents the total number of real label samples from the high-confidence honeypot extracted each time. This indicates the number of samples whose pseudo-labels match the true labels of the high-confidence honeypot. Indicates the accuracy of pseudo-labels. This indicates the number of samples whose pseudo-labels changed between two adjacent iterations. This indicates the total number of samples involved in pseudo-label generation in the current round. This indicates consistency across rounds.
[0009] Preferably, the feature discrimination calculation unit calculates the average discrimination of multi-dimensional features. The calculation formula is as follows: ; In the formula, The average discriminant value representing multi-dimensional features. This represents the total dimension of the features involved in the calculation. Indicates the first An index of features to be evaluated. Indicates the first The set of values for each feature Binary tags representing traffic. Indicates the first Each feature is taken And the traffic tag is The joint probability, Indicates the first Each feature is taken The marginal probability, Indicates traffic label as The marginal probability.
[0010] Preferably, the adversarial perturbation threshold determination unit calculates the characteristic perturbation threshold for adversarial attack determination. The calculation formula is as follows: ;
[0011] In the formula, The threshold for feature perturbation used to determine adversarial attacks. This represents the mean of characteristic perturbations in a normal flow sample. This represents the characteristic fluctuation of the j-th normal sample. This represents the total number of normal traffic samples used for calculation. It represents the standard deviation of the characteristic perturbation of normal flow samples.
[0012] Preferably, the annotation quality verification module scores the annotation data based on its quality. The process involves comparing pseudo-labels with real labels and verifying label consistency across iteration rounds. When labeling data quality score When the value is ≥0.8 of the preset qualified threshold, the pseudo-label quality is confirmed to be qualified and can be used for model training; When labeling data quality score When the preset qualified threshold is 0.8, the data is fed back to the intelligent data acquisition module to regenerate the pseudo-label.
[0013] Preferably, the feature validity determination module is based on the average discriminative power of multi-dimensional features. Cross-protocol feature discrimination comparison and feature-tag mutual information verification are performed, specifically as follows: When the average discriminative power of multidimensional features If the preset validity threshold is ≥0.3, the feature set is confirmed to be valid and can be input into the detection model. When the average discriminative power of multidimensional features When the preset validity threshold is 0.3, the feature extraction and filtering module is fed back to re-filter / map the features.
[0014] Preferably, the adversarial perturbation detection module uses a characteristic perturbation threshold determined by the adversarial attack assessment. Calculate the characteristic perturbations of real-time traffic samples and the characteristic perturbation threshold for adversarial attack determination The comparison is as follows: When the characteristics of real-time traffic samples are perturbed ≤ Feature perturbation threshold for adversarial attack determination If the sample is identified as a normal flow sample, it will proceed to the next testing process. When the characteristics of real-time traffic samples are perturbed >Characteristic perturbation threshold for adversarial attack detection When this occurs, it is marked as an adversarial example and a defense mechanism is triggered.
[0015] Compared with existing technologies, this invention provides a deep learning-based encrypted malicious traffic detection system, which has the following beneficial effects: 1. This invention calculates the quality score of labeled data. This method is used to evaluate the accuracy and stability of pseudo-labels generated by semi-supervised learning. It is applied to the label quality verification module as a criterion for judging whether pseudo-labels are qualified, ensuring the quality of labeled data used for model training, reducing the negative impact of low-quality pseudo-labels on model performance, and improving the model's generalization ability in new attack scenarios.
[0016] 2. This invention utilizes the average discriminative power of the aforementioned multi-dimensional features. The calculation is used as a quantitative indicator of feature effectiveness, which is then used in the feature effectiveness judgment module to screen high-discrimination features. This ensures that the features input to the detection model can effectively characterize the differences between malicious and normal traffic, improve the model's ability to identify and distinguish features of encrypted traffic, and enhance cross-protocol transferability.
[0017] This invention calculates the feature perturbation threshold for adversarial attack determination. This is used as a criterion for judging adversarial examples, and is used in the adversarial perturbation detection module to identify adversarial examples with small perturbations, thereby improving the system's ability to defend against adversarial attacks, reducing the risk of attackers misleading the model by generating adversarial examples, and ensuring the stability of the detection system in adversarial scenarios. Attached Figure Description
[0018] Figure 1 This is a system flowchart of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Please see Figure 1 A deep learning-based encrypted malicious traffic detection system includes an intelligent data acquisition module, a feature extraction and filtering module, an intelligent data analysis module, an annotation quality verification module, a metadata enhancement detection module, an adversarial disturbance detection module, a dynamic collaborative control module, a real-time online feedback module, and a full-scenario validity verification module. The intelligent data acquisition module is responsible for collecting raw metadata from multiple sources automatically, and alleviates the problem of scarce labeled data through technical means; The feature extraction and filtering module constructs high-dimensional features from the original metadata, filters noise, and extracts robust feature representations across protocols and environments. The intelligent data analysis module receives feature input from the feature extraction and filtering module and calculates the quality score of the labeled data. Average discrimination of multi-dimensional features Feature perturbation threshold for adversarial attack determination This quantifies the risks and uncertainties of current traffic across different dimensions, providing a basis for decision-making in subsequent specialized modules. The annotation quality verification module is based on the quality score of the annotation data. The calculation formula is used to compare pseudo-labels with real labels and to verify the consistency of labels across iteration rounds. The feature validity determination module is based on the average discriminative power of multi-dimensional features. The calculation formula is used to compare cross-protocol feature distinguishability and verify feature-tag mutual information. The adversarial perturbation detection module uses feature perturbation thresholds based on adversarial attack determination. The calculation formula is used to calculate the feature perturbation of the current traffic sample and compare it with the threshold. Based on the output results of the annotation quality verification module, feature validity judgment module, and adversarial disturbance detection module, the dynamic collaborative control module generates control commands and issues them to the intelligent data acquisition module and feature extraction and filtering module, thereby solving the problem of poor adaptability to dynamic network environments and realizing system parameter self-adaptation. The real-time online feedback module monitors the operating status and output indicators of each module (such as fluctuations in annotation quality scores, changes in feature discrimination, and the number of adversarial sample detections) online through a real-time data stream monitoring and indicator feedback mechanism. It captures abnormal information in real time and updates strategies online, solving the problems of no universal solution for defense measures and difficulty in timely handling of vulnerabilities, and ensuring the real-time performance of the system. The full-scenario validity verification module performs full-process testing on the system's annotation quality, feature validity, adversarial defense capabilities, and dynamic adaptation capabilities in various encryption scenarios (different TLS versions, IoT sparse traffic, new encryption algorithm scenarios, etc.), verifying whether all problems have been resolved and ensuring system availability.
[0021] The advantages are: through the above-mentioned multi-module collaboration and quantitative decision-making mechanism, the four core problems faced in encrypted traffic detection are systematically solved, namely, data scarcity, limited features, adversarial attacks and environment adaptation, and the whole process optimization and self-adaptation from the data source to the final decision are realized.
[0022] The intelligent data acquisition module includes a metadata acquisition unit, a semi-supervised annotation unit, and an adversarial enhancement unit; The metadata collection unit extracts metadata such as packet length, timing, direction, and plaintext information of the TLS handshake from the traffic. Semi-supervised labeling units generate pseudo-labels for unlabeled data based on pre-trained teacher models; The adversarial enhancement unit applies controlled adversarial perturbations to existing labeled data to generate adversarial examples for training.
[0023] The advantages are: by combining the above-mentioned multi-dimensional collection of metadata, semi-supervised pseudo-label generation and adversarial example enhancement, the problem of scarce labeled data is alleviated, while enriching the diversity of training data (covering normal and adversarial scenarios), providing a high-quality and high-coverage training foundation for subsequent detection models.
[0024] The feature extraction and filtering module includes a spatiotemporal feature construction unit and a noise filtering unit; The spatiotemporal feature construction unit transforms the packet-length sequence and interaction interval into feature vectors; The noise filtering unit filters low-value or high-variance features based on feature importance scores.
[0025] The advantages are: through the above spatiotemporal feature construction, noise filtering and protocol-independent encoding, the original metadata is transformed into a robust feature representation that is cross-protocol and noise-resistant, which improves the feature's ability to characterize malicious traffic patterns, provides clean and effective features for the detection model, and reduces the interference of irrelevant noise on detection accuracy.
[0026] The intelligent data analysis module includes a labeled data quality assessment unit (to assess the accuracy of pseudo-labels), a feature discrimination calculation unit (to assess the effectiveness of features), and an adversarial perturbation threshold determination unit (to determine the threshold for adversarial attack judgment).
[0027] The annotation data quality assessment unit calculates the annotation data quality score. The calculation formula is as follows: ; In the formula, This indicates the quality score of the labeled data. The weight representing the accuracy of pseudo-labels (used to adjust the influence of the accuracy of matching pseudo-labels with real labels on the overall score). The weight representing the consistency of pseudo-labels across iterations (used to adjust the impact of the stability of pseudo-labels across different iterations on the overall score). This represents the total number of real label samples from the high-confidence honeypot extracted each time (i.e., the total number of real label samples extracted from the honeypot system to verify the accuracy of the fake labels, which is set to 200 in this paper). This indicates the number of samples whose pseudo-labels match the true labels of the high-confidence honeypot. This represents the accuracy of the pseudo-label (the correct matching rate between the pseudo-label and the real label of the high-confidence honeypot). This indicates the number of samples whose pseudo-labels changed between two adjacent iterations. This indicates the total number of samples involved in pseudo-label generation in the current round. This indicates cross-cycle consistency (the stability of pseudo-labels across adjacent iterations; the fewer the number of changing samples, the higher the cross-cycle consistency).
[0028] The advantage is that it calculates the quality score of the labeled data. This method is used to evaluate the accuracy and stability of pseudo-labels generated by semi-supervised learning. It is applied to the label quality verification module as a criterion for judging whether pseudo-labels are qualified, ensuring the quality of labeled data used for model training, reducing the negative impact of low-quality pseudo-labels on model performance, and improving the model's generalization ability in new attack scenarios.
[0029] The feature discrimination calculation unit (evaluating feature effectiveness) calculates the average discrimination of multi-dimensional features. The calculation formula is as follows: ; In the formula, The average discriminant value representing multi-dimensional features. This indicates the total number of dimensions of features involved in the computation (covering metadata features, plaintext parameter features during the cryptographic handshake phase, traffic timing features, etc.). Indicates the first Index of each feature to be evaluated (traversing all) (characteristics) Indicates the first The set of possible values for each feature (such as possible values for packet length, optional parameters of TLS cipher suites, etc.). The traffic is represented by a binary label (malicious traffic is labeled as 1, and normal traffic is labeled as 0). Indicates the first Each feature is taken And the traffic tag is The joint probability, Indicates the first Each feature is taken Marginal probabilities (without considering labels, features) (probability of occurrence) Indicates traffic label as Marginal probabilities (without considering features, label) (probability of occurrence).
[0030] The advantage is that it achieves the average discriminative power of the aforementioned multi-dimensional features. The calculation is used as a quantitative indicator of feature effectiveness, which is then used in the feature effectiveness judgment module to screen high-discrimination features. This ensures that the features input to the detection model can effectively characterize the differences between malicious and normal traffic, improve the model's ability to identify and distinguish features of encrypted traffic, and enhance cross-protocol transferability.
[0031] The adversarial perturbation threshold determination unit (determines the adversarial attack determination threshold) calculates the characteristic perturbation threshold for adversarial attack determination. The calculation formula is as follows: ; In the formula, The feature perturbation threshold for adversarial attack determination (when the feature perturbation of a traffic sample exceeds this threshold, it is determined to be an adversarial sample). This represents the mean of characteristic perturbations in a normal flow sample. This represents the characteristic fluctuation of the j-th normal sample (i.e., the natural fluctuation of the normal flow sample in the feature space, which is calculated from the difference between the normal sample and the baseline characteristics when there is no disturbance). This represents the total number of normal traffic samples used for calculation. It represents the standard deviation of the characteristic perturbation of normal flow samples.
[0032] The advantage is that it calculates the feature perturbation threshold for adversarial attack detection. This is used as a criterion for judging adversarial examples, and is used in the adversarial perturbation detection module to identify adversarial examples with small perturbations, thereby improving the system's ability to defend against adversarial attacks, reducing the risk of attackers misleading the model by generating adversarial examples, and ensuring the stability of the detection system in adversarial scenarios.
[0033] The annotation quality verification module scores the annotation data based on its quality. Perform comparison between pseudo-labels and real labels, and verify label consistency across iterations; when the labeled data quality score is... When the pseudo-label quality score is ≥ 0.8 (preset acceptable threshold), the pseudo-label quality is considered acceptable and can be used for model training; when the labeled data quality score is... When the preset qualified threshold is 0.8, the data is fed back to the intelligent data acquisition module to regenerate the pseudo-label.
[0034] The advantages are: by using the pseudo-label quality scoring mechanism to determine whether the pseudo-labels are qualified or unqualified, a closed-loop control of pseudo-label quality is achieved, ensuring that the labeled data used for model training always maintains high quality, providing a guarantee for the performance of the detection model from the data level, and alleviating the problems of lag and high cost of manual annotation.
[0035] The feature validity determination module determines the effectiveness based on the average discriminative power of multi-dimensional features. Perform cross-protocol feature discrimination comparison and feature-label mutual information verification; when the average discrimination of multi-dimensional features... When the average discrimination of the multi-dimensional features is greater than or equal to the preset validity threshold of 0.3, the feature set is confirmed to be valid and allowed to be input into the detection model; when the average discrimination of the multi-dimensional features is... When the preset validity threshold is 0.3, the feature extraction and filtering module is fed back to re-filter / map the features.
[0036] The advantages are: by using a feature validity / invalidity determination and feedback mechanism based on the average discriminative power of multi-dimensional features, closed-loop control of feature quality is achieved, ensuring that the features input to the detection model have strong discriminative power, reducing the interference of redundant metadata and noisy features on the model, and improving the detection accuracy of the model under different encryption protocols and dynamic environments.
[0037] The adversarial perturbation detection module determines the characteristic perturbation threshold based on adversarial attacks. Calculate the characteristic perturbations of real-time traffic samples and the characteristic perturbation threshold for adversarial attack determination Comparison; when the characteristics of real-time traffic samples are perturbed ≤ Feature perturbation threshold for adversarial attack determination When the real-time flow sample is disturbed, it is determined to be a normal flow sample and proceeds to the subsequent detection process; when the characteristics of the real-time flow sample are disturbed... >Characteristic perturbation threshold for adversarial attack detection When this occurs, it is marked as an adversarial example and a defense mechanism is triggered.
[0038] The advantages are: by using a real-time detection and defense triggering mechanism for adversarial samples based on feature perturbation thresholds, timely identification and response to adversarial attacks can be achieved, making up for the vulnerability of deep learning models to adversarial attacks, ensuring that the system can still accurately detect malicious traffic when facing adversarial attacks, and improving the security in actual deployment scenarios.
[0039] In summary, this invention addresses the challenges of obtaining labeled data, the susceptibility of features to interference, vulnerability to adversarial attacks, and poor adaptability to dynamic environments in encrypted malicious traffic detection. It constructs a multi-module collaborative deep learning detection system. The system employs a comprehensive design that addresses the scarcity of labeled data through intelligent data acquisition, ensures feature quality through feature extraction and filtering, quantifies core indicators through intelligent data analysis, and achieves closed-loop control and defense through multiple specialized modules. Ultimately, this system enables accurate, real-time, and robust detection of malicious traffic in encrypted scenarios, providing an efficient solution for combating encrypted malicious traffic in the cybersecurity field.
[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A deep learning-based encrypted malicious traffic detection system, characterized in that, It includes an intelligent data acquisition module, a feature extraction and filtering module, an intelligent data analysis module, an annotation quality verification module, a metadata enhancement detection module, an adversarial disturbance detection module, a dynamic collaborative control module, a real-time online feedback module, and a full-scenario validity verification module; The intelligent data acquisition module is responsible for collecting raw metadata; The feature extraction and filtering module constructs high-dimensional features from the original metadata, filters noise, and extracts robust feature representations across protocols and environments. The intelligent data analysis module receives feature input from the feature extraction and filtering module and calculates the labeled data quality score. Average discrimination of multi-dimensional features Feature perturbation threshold for adversarial attack determination ; The annotation quality verification module is based on annotation data quality scoring. The calculation formula is used to compare pseudo-labels with real labels and to verify the consistency of labels across iteration rounds. The feature validity determination module is based on the average discrimination of multi-dimensional features. The calculation formula is used to compare cross-protocol feature distinguishability and verify feature-tag mutual information. The adversarial disturbance detection module is based on the characteristic disturbance threshold for adversarial attack determination. The calculation formula is used to calculate the feature perturbation of the current traffic sample and compare it with the threshold. The dynamic collaborative control module generates control commands based on the output results of the annotation quality verification module, feature validity determination module, and anti-disturbance detection module, and issues them to the intelligent data acquisition module and feature extraction and filtering module to achieve adaptive system parameters. The real-time online feedback module monitors the operating status and output indicators of each module online through a real-time data stream monitoring and indicator feedback mechanism, and performs real-time capture of abnormal information and online updating of strategies. The full-scenario validity verification module performs full-process testing on the system's annotation quality, feature validity, adversarial defense capabilities, and dynamic adaptation capabilities under multiple encryption scenarios, verifying whether all issues have been resolved.
2. The encrypted malicious traffic detection system based on deep learning according to claim 1, characterized in that, The intelligent data acquisition module includes a metadata acquisition unit, a semi-supervised annotation unit, and an adversarial enhancement unit; The metadata collection unit extracts metadata such as packet length, timing, direction, and plaintext information of the TLS handshake from the traffic. The semi-supervised labeling unit generates pseudo-labels for unlabeled data based on a pre-trained teacher model. The adversarial enhancement unit applies controllable adversarial perturbations to existing labeled data to generate adversarial samples for training.
3. The encrypted malicious traffic detection system based on deep learning according to claim 1, characterized in that, The feature extraction and filtering module includes a spatiotemporal feature construction unit and a noise filtering unit; The spatiotemporal feature construction unit converts the packet-length sequence and interaction interval into feature vectors; The noise filtering unit filters low-value or high-variance features based on feature importance scores.
4. The encrypted malicious traffic detection system based on deep learning according to claim 1, characterized in that, The intelligent data analysis module includes a labeled data quality assessment unit, a feature discrimination calculation unit, and an adversarial perturbation threshold determination unit.
5. The encrypted malicious traffic detection system based on deep learning according to claim 4, characterized in that, The annotation data quality assessment unit calculates the annotation data quality score. The calculation formula is as follows: ； In the formula, This indicates the quality score of the labeled data. The weights representing the accuracy of pseudo-labels The weight representing the consistency of pseudo-labels across rounds. This represents the total number of real label samples from the high-confidence honeypot extracted each time. This indicates the number of samples whose pseudo-labels match the true labels of the high-confidence honeypot. Indicates the accuracy of pseudo-labels. This indicates the number of samples whose pseudo-labels changed between two adjacent iterations. This indicates the total number of samples involved in pseudo-label generation in the current round. This indicates consistency across rounds.
6. The encrypted malicious traffic detection system based on deep learning according to claim 4, characterized in that, The feature discrimination calculation unit calculates the average discrimination of multi-dimensional features. The calculation formula is as follows: ； In the formula, The average discriminant value representing multi-dimensional features. This represents the total dimension of the features involved in the calculation. Indicates the first An index of features to be evaluated. Indicates the first The set of values for each feature Binary tags representing traffic. Indicates the first Each feature is taken And the traffic tag is The joint probability, Indicates the first Each feature is taken The marginal probability, Indicates traffic label as The marginal probability.
7. The encrypted malicious traffic detection system based on deep learning according to claim 4, characterized in that, The anti-disturbance threshold determination unit calculates the characteristic perturbation threshold for anti-attack determination. The calculation formula is as follows: ； In the formula, The threshold for feature perturbation used to determine adversarial attacks. This represents the mean of characteristic perturbations in a normal flow sample. This represents the characteristic fluctuation of the j-th normal sample. This represents the total number of normal traffic samples used for calculation. It represents the standard deviation of the characteristic perturbation of normal flow samples.
8. The encrypted malicious traffic detection system based on deep learning according to claim 1, characterized in that, The annotation quality verification module scores the annotation data quality. The process involves comparing pseudo-labels with real labels and verifying label consistency across iteration rounds. When labeling data quality score When the value is ≥0.8 of the preset qualified threshold, the quality of the pseudo-labels is confirmed to be qualified and can be used for model training; When labeling data quality score When the preset qualified threshold is 0.8, the data is fed back to the intelligent data acquisition module to regenerate the pseudo-label.
9. The encrypted malicious traffic detection system based on deep learning according to claim 1, characterized in that, The feature validity determination module is based on the average discrimination of multi-dimensional features. Cross-protocol feature discrimination comparison and feature-tag mutual information verification are performed, specifically as follows: When the average discriminative power of multidimensional features If the preset validity threshold is ≥0.3, the feature set is confirmed to be valid and can be input into the detection model. When the average discriminative power of multidimensional features When the preset validity threshold is 0.3, the feature extraction and filtering module is fed back to re-filter / map the features.
10. A deep learning-based encrypted malicious traffic detection system according to claim 1, characterized in that, The anti-disturbance detection module determines the characteristic disturbance threshold based on the anti-attack criteria. Calculate the characteristic perturbations of real-time traffic samples and the characteristic perturbation threshold for adversarial attack determination The comparison is as follows: When the characteristics of real-time traffic samples are perturbed ≤ Feature perturbation threshold for adversarial attack determination If the sample is identified as a normal flow sample, it will proceed to the next testing process. When the characteristics of real-time traffic samples are perturbed >Characteristic perturbation threshold for adversarial attack detection When this occurs, it is marked as an adversarial example and a defense mechanism is triggered.