A safety operation method and system based on multi-agent collaboration

Abstract

This invention relates to the field of data security technology, specifically to a security operation method and system based on multi-agent collaboration. The method includes: utilizing an edge-aware agent to intercept the TLS length and time interval sequence of encrypted sessions; constructing a spatiotemporal topology embedding vector based on linear visibility criteria and graph attention mechanism; mapping this vector to Gaussian distribution parameters to generate token adjustment coefficients; dynamically adjusting the token bucket frequency and performing asynchronous delay shaping using a time wheel; and a reflective agent combining traffic obfuscation escape rate and processing latency indicators to perform closed-loop updates of model parameters using a policy gradient algorithm. This invention achieves adaptive defense of encrypted covert channels through dynamic traffic shaping.

Description

Technical Field

[0001] This invention relates to the field of data security technology, specifically to a security operation method and system based on multi-agent collaboration. Background Technology

[0002] With the deepening of digital transformation, enterprise data services and the construction of urban public data spaces are becoming increasingly popular. Application programming interfaces (APIs) have become the main channel for data interaction. In order to ensure the confidentiality and integrity of data transmission, transport layer security protocols (such as TLS / SSL) are widely used in various business scenarios, and encrypted communication has become the norm for data services.

[0003] However, while encryption technology protects data content from being stolen, it also provides a natural cover for malicious traffic. In actual security operations, attackers or malicious programs often no longer send attack payloads directly in plaintext, but instead use encrypted sessions to establish covert channels. These covert channels usually do not rely on the effective payload content of data packets, but transmit information by manipulating the non-semantic physical characteristics of the data stream. For example, they use side-channel characteristics such as changes in the length sequence of data packets, the time interval pattern of message arrival, and the burst rhythm of traffic to steal data or control commands.

[0004] Existing security systems typically rely on deep inspection of decrypted content or blocking based on fixed traffic statistics thresholds. Faced with complex, covert channels built using encryption protocols, static rule matching or single-dimensional statistical analysis often fails to accurately identify low-frequency, slow transmission behaviors that simulate normal business operations. Conversely, overly aggressive blocking strategies can easily inadvertently damage normal business data transmission, leading to increased service latency or interruptions.

[0005] Therefore, how to accurately perceive the abnormal transmission timing characteristics in encrypted sessions without decrypting the data content, and to dynamically and adaptively regulate and defend against potential covert transmission behaviors, while balancing the effectiveness of security protection and the quality of business services, is a technical problem that urgently needs to be solved in the field of data security operations.

[0006] To address this, a safe operation method and system based on multi-agent collaboration is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a security operation method and system based on multi-agent collaboration, which achieves adaptive defense of encrypted covert channels through dynamic traffic shaping. This includes using an edge-aware agent to intercept the TLS length and time interval sequence of encrypted sessions; constructing a spatiotemporal topology embedding vector based on the linear visibility criterion and graph attention mechanism; mapping this vector to Gaussian distribution parameters to generate token adjustment coefficients; dynamically adjusting the token bucket frequency and performing asynchronous delay shaping using a time wheel; and using a reflective agent to update model parameters in a closed loop by combining traffic obfuscation escape rate and processing latency indicators through a policy gradient algorithm.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A security operation method based on multi-agent collaboration includes:

[0010] The edge-aware agent intercepts the TLS record layer length sequence and message arrival time interval sequence of the encrypted session from the data service gateway; it maps the time interval sequence into a temporal topological adjacency matrix based on the linear visibility geometry criterion, and uses an attention mechanism to weight the length sequence and the matrix to generate a spatiotemporal topological embedding vector.

[0011] The embedded vector input strategy model is mapped to multidimensional Gaussian distribution parameters. The traffic agent samples the distribution parameters to generate token adjustment coefficients. The token distribution frequency of the gateway token bucket algorithm is updated using the adjustment coefficients. The target release timestamp of the data packets to be sent is calculated based on the current token status in the bucket. Asynchronous delay is performed on data packets that have not reached the timestamp. The response traffic flow is output, and the gateway processing delay index introduced by the asynchronous delay is recorded.

[0012] The red team agent uses random forest to perform business fingerprinting on the response traffic flow and generates traffic confusion escape rate; the reflective agent reads latency indicators and escape rate to calculate the generalized advantage function estimate and the action probability ratio of the old and new strategies, constructs the substitution objective function using a preset cutoff threshold, and adjusts the parameters of the strategy model through the stochastic gradient ascent algorithm to maximize the substitution objective function.

[0013] Preferably, the process by which the edge-aware agent intercepts data from the data service gateway includes: the edge-aware agent is constructed as an independent running instance containing a traffic capture probe module and a time-series feature preprocessing module, and deployed on the network-side channel observation interface of the data service gateway; the traffic capture probe module reads the binary data frames in the gateway network buffer in real time, locates the secure transport layer protocol record layer header structure therein, and parses the content type field in the header structure; only when the content type field identifies the current data frame as an application layer data protocol, the payload length field value immediately following the content type field is extracted, normalized and stored in the TLS record layer length sequence, and control data frames of handshake negotiation protocol and alarm protocol type are filtered and discarded; the time-series feature preprocessing module reads the hardware receive timestamp of the data frame, calculates the first-order difference between the timestamp of the current data frame and the timestamp of the previous data frame to obtain the original interval data, and performs natural logarithmic transformation and Z-Score normalization on the original interval data to eliminate the influence of long-tail distribution and generate the message arrival time interval sequence.

[0014] Preferably, the process of generating the spatiotemporal topology embedding vector includes: the edge-aware agent establishing a temporal-numerical two-dimensional coordinate system, mapping the message arrival time interval sequence to a set of nodes in the coordinate system, where the horizontal axis is the message temporal index and the vertical axis is the time interval value; for any two target nodes in the node set, calculating the linear slope of the line connecting them; if the vertical axis values ​​of all intermediate nodes are lower than the linear interpolation height of the two target nodes at their corresponding temporal positions, then it is determined that the two target nodes have a visible connection relationship, and the corresponding element in the temporal topology adjacency matrix is ​​set to a non-zero connection. The process involves: identifying and constructing a graph attention network with multiple attention heads using an edge-aware agent; mapping the TLS record layer length sequence to high-dimensional node feature vectors using a linear projection matrix; calculating the dot product correlation coefficient between adjacent node feature vectors and using a temporal topological adjacency matrix as a structural mask to set the correlation coefficient of non-connected nodes to negative infinity; normalizing the correlation coefficients after masking to generate an attention weight distribution and weighting and aggregating the node feature vectors in the neighborhood; and concatenating the output vectors of multiple attention heads and generating a spatiotemporal topological embedding vector through a fully connected layer.

[0015] Preferably, the edge-aware agent also intercepts the gateway's traffic throughput differential sequence, token bucket remaining capacity, and system load level; the strategy model includes:

[0016] The topology encoder receives the spatiotemporal topology embedding vector as input, uses a residual fully connected network for nonlinear mapping processing, extracts the high-dimensional structured dependencies hidden in the encrypted traffic, and obtains a static topology feature vector.

[0017] The differential trend encoder receives the traffic throughput difference sequence between the current time window and the previous time window of the gateway as input, and uses a one-dimensional convolutional neural network to extract time-series features and generate a dynamic trend feature vector.

[0018] The global state encoder receives the current token bucket remaining capacity and system load level of the gateway as input, and processes them using a preset state embedding layer and a multilayer perceptron to map the resource constraint state of the physical layer into a global constraint feature vector.

[0019] The multimodal gating fusion unit performs a channel-dimensional concatenation operation on the dynamic trend feature vector and the global constraint feature vector to generate an environmental context vector; inputs the environmental context vector into the attention generation network, and uses the Sigmoid activation function to generate a feature recalibration weight vector with a value range of 0 to 1; performs an element-wise Hadamard product operation on the feature recalibration weight vector and the static topological feature vector to output a global content-aware state vector;

[0020] The distribution parameter mapping layer receives the global content-aware state vector and constructs a parallel mean generation linear layer and a standard deviation generation linear layer. The mean generation linear layer maps the input vector to an action mean vector. The standard deviation generation linear layer maps the input vector to an action logarithmic standard deviation vector and performs a natural exponential transformation. The action mean vector and the transformed standard deviation vector are combined to output a multidimensional Gaussian distribution parameter.

[0021] Preferably, the process of outputting the response traffic stream and recording the gateway processing latency metric introduced by asynchronous latency includes: the traffic agent employs a reparameterization technique to sample random noise from a standard normal distribution and calculates the token adjustment coefficient by combining it with the distribution parameters output by the policy model; it reads the gateway's preset baseline token delivery rate, multiplies the baseline token delivery rate with the token adjustment coefficient to generate the dynamic filling rate within the current time window; it obtains the payload bytes of the data packet to be sent and queries the remaining token count in the current token bucket; when the payload bytes are greater than the remaining token count, it calculates the difference between the two. The difference is divided by the dynamic filling rate to calculate the waiting time required for the data packet to acquire a token equal to the number of bytes in the data packet payload; the current system kernel time is read and added to the waiting time to determine the target release timestamp of the data packet; a delayed task object containing a data packet reference and a callback function is constructed, and the task object is registered to the corresponding time slot of the hierarchical time wheel scheduler according to the target release timestamp; when the system time advances to the target release timestamp, the scheduler triggers the callback function, writes the data packet into the gateway sending queue, and accumulates the waiting time into the statistical buffer of the gateway processing latency index.

[0022] Preferably, the backpropagation update of the policy model parameters using the stochastic gradient ascent algorithm includes: adjusting the policy model parameters using the stochastic gradient ascent algorithm to maximize the substitution objective function includes: the red team agent extracting statistical feature vectors of the response traffic flow and inputting them into a preset random forest classifier; the classifier contains multiple independently trained decision trees, each decision tree outputting a business category prediction result; the output results of all decision trees are statistically analyzed, and the voting percentage of the correct business label is calculated as the recognition confidence; the difference between 1 and the recognition confidence is calculated, and the difference is determined as the traffic confusion escape rate; the reflective agent uses a weighted linear combination method and a weighted calculation method to map the normalized traffic confusion escape rate as a positive reward term and the normalized gateway processing latency indicator as a negative penalty term, and calculates the difference between the positive reward term and the negative penalty term, and determines the difference as the instant reward value at the current moment; using a preset The value network predicts the state value scalars of the current state and the state at the next time step; it calculates the temporal difference error by combining the immediate reward value and the state value scalar; it introduces a discount factor and a smoothing parameter, and performs an exponentially weighted moving average calculation on multiple temporal difference errors within a future time window to obtain the generalized dominance function estimate; it reads the parameters of the policy model before and after the update, calculates the probability of the new policy and the probability of the old policy to perform the same action, and calculates the ratio between the two; it calculates the product of the ratio and the generalized dominance function estimate to obtain the untruncated target term; simultaneously, it restricts the ratio to a preset floating range centered at 1, calculates the product of the restricted ratio and the generalized dominance function estimate to obtain the truncated target term; it selects the minimum value between the untruncated target term and the truncated target term as the alternative target function, uses the stochastic gradient ascent algorithm to calculate the gradient of the alternative target function with respect to the policy model parameters, and performs backpropagation to update the policy model parameters.

[0023] A security operation system based on multi-agent collaboration includes:

[0024] The edge perception module intercepts the TLS record layer length sequence and message arrival time interval sequence of the encrypted session from the data service gateway; it maps the time interval sequence into a temporal topological adjacency matrix according to the linear visibility geometry criterion, and uses an attention mechanism to weight the length sequence and the matrix to generate a spatiotemporal topological embedding vector.

[0025] The traffic shaping module maps the embedded vector input strategy model to multidimensional Gaussian distribution parameters, uses the traffic agent to sample and generate token adjustment coefficients, uses the adjustment coefficients to update the token delivery frequency of the gateway token bucket algorithm, calculates the target release timestamp of the data packets to be sent based on the current token status in the bucket, performs asynchronous delay operations on data packets that have not reached the timestamp, outputs the response traffic stream, and records the gateway processing delay index introduced by the asynchronous delay.

[0026] The adversarial assessment module uses a random forest classifier to perform business fingerprinting on the response traffic flow and generate a traffic obfuscation escape rate.

[0027] The optimization module reads latency metrics and escape rates to calculate the generalized advantage function estimate and the action probability ratio between the old and new strategies. It then constructs an alternative objective function using a preset cutoff threshold and adjusts the policy model parameters using a stochastic gradient ascent algorithm to maximize the alternative objective function.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] 1. By constructing a spatiotemporal topology embedding vector using linear visibility geometry criteria and graph attention mechanism, the message arrival time interval sequence of encrypted sessions is mapped into a temporal topology adjacency matrix with complex network characteristics. Furthermore, the TLS record layer length attribute is integrated, thereby accurately extracting high-dimensional structured dependencies hidden in the physical side channel without decrypting the service payload. This method effectively identifies covert channels constructed using low-frequency, slow, or mimicry techniques, significantly improving the data service gateway's ability to perceive and detect unknown encryption threats.

[0030] 2. By combining multidimensional Gaussian distribution sampling with the inverse token bucket algorithm to perform dynamic traffic shaping, and utilizing reparameterization techniques to generate token adjustment coefficients with random exploratory capabilities, a hierarchical time-wheel scheduler performs microsecond-level asynchronous delay operations on data packets, thereby reconstructing the temporal characteristics of traffic in real time while avoiding data loss. This proactive interference mechanism effectively smooths out the physical fingerprint differences required for attacker side-channel analysis, while the efficient asynchronous scheduling architecture avoids thread blocking in high-concurrency scenarios, ensuring that the gateway maintains high data throughput performance even when executing complex shaping strategies.

[0031] 3. By constructing a closed-loop feedback mechanism based on red team adversarial evaluation and policy gradient optimization, the agent reflects on how to use the generalized advantage function to estimate and truncate the update policy. While pursuing high security gains, the gateway processing latency is included in the negative penalty, which effectively prevents network service jitter caused by excessive policy update or over-defense. This self-game mechanism enables the defense policy to continuously evolve in the dynamically changing network adversarial environment, and finally converges to the optimal state where security confusion effect and business service quality are balanced, thus enhancing the system's adaptability.

[0032] 4. By designing a strategy model based on multi-view feature fusion and gating mechanisms, and utilizing the feature recalibration weights generated by the multimodal gating fusion unit, the system can dynamically suppress topological feature channels affected by noise based on the current traffic fluctuation trend and resource load level, enhancing the model's robustness under network jitter or sudden traffic conditions. Combined with a heterogeneous coding architecture of residual fully connected networks and one-dimensional convolutional neural networks, it effectively solves the problem of insufficient information representation from a single feature perspective in complex attack and defense scenarios, improving the agent's decision-making generalization ability in different network environments. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of a security operation method based on multi-agent collaboration according to the present invention;

[0034] Figure 2 This is a schematic diagram of the process for generating spatiotemporal topology embedding vectors according to the present invention;

[0035] Figure 3 This is a schematic diagram of a security operation system structure based on multi-agent collaboration according to the present invention. Detailed Implementation

[0036] 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.

[0037] Please see Figures 1 to 3 This invention provides a safe operation method and system based on multi-agent collaboration, the technical solution of which is as follows:

[0038] Example 1:

[0039] A security operation method based on multi-agent collaboration, the specific process is as follows: Figure 1 The following are included:

[0040] The edge-aware agent intercepts the TLS record layer length sequence and message arrival time interval sequence of the encrypted session from the data service gateway; it maps the time interval sequence into a temporal topological adjacency matrix based on the linear visibility geometry criterion, and uses an attention mechanism to weight the length sequence and the matrix to generate a spatiotemporal topological embedding vector.

[0041] The embedded vector input strategy model is mapped to multidimensional Gaussian distribution parameters. The traffic agent samples the distribution parameters to generate token adjustment coefficients. The token distribution frequency of the gateway token bucket algorithm is updated using the adjustment coefficients. The target release timestamp of the data packets to be sent is calculated based on the current token status in the bucket. Asynchronous delay is performed on data packets that have not reached the timestamp. The response traffic flow is output, and the gateway processing delay index introduced by the asynchronous delay is recorded.

[0042] The red team agent uses random forest to perform business fingerprinting on the response traffic flow and generates traffic confusion escape rate; the reflective agent reads latency indicators and escape rate to calculate the generalized advantage function estimate and the action probability ratio of the old and new strategies, constructs the substitution objective function using a preset cutoff threshold, and adjusts the parameters of the strategy model through the stochastic gradient ascent algorithm to maximize the substitution objective function.

[0043] In this embodiment, multiple agents interact with each other through a shared state buffer. Specifically, the state information, such as the spatiotemporal topology embedding vector and traffic throughput differential sequence calculated by the edge-aware agent, is written into a shared circular buffer. The traffic agent reads the latest state information from this buffer every fixed time interval (50 milliseconds in this embodiment) and performs an inference calculation to generate a token adjustment coefficient, thereby ensuring that the latency in response to network changes is controlled within a predetermined range. The response traffic stream output by the traffic agent is mirrored to the input buffer of the red team agent. The red team agent performs statistical feature extraction and classification on the traffic, calculates the traffic confusion escape rate, and synchronizes the results. The reward information is written into a shared buffer for consumption by the reflexive agent. After the reflexive agent completes a gradient update of the policy parameters once every learning cycle (e.g., 1 second), it uses an atomic parameter exchange operation (synchronization is ensured by locking) to copy the new policy parameters from the reflexive agent's parameter storage area to the active inference engine of the traffic agent. At the same time, the traffic agent immediately stops using the old parameters and starts using the new parameters to process subsequent data packets. Data packets before the parameter exchange is completed continue to use the old parameters, ensuring the smoothness of the transition. Through this design, multiple agents can work efficiently in a loosely coupled architecture while avoiding the blocking problem caused by synchronous communication.

[0044] Furthermore, the process by which the edge-aware intelligent agent intercepts data from the data service gateway includes: the edge-aware intelligent agent is constructed as an independent running instance containing a traffic capture probe module and a time-series feature preprocessing module, and deployed on the network side channel observation interface of the data service gateway; the traffic capture probe module reads the binary data frames in the gateway network buffer in real time, locates the secure transport layer protocol record layer header structure, and parses the content type field in the header structure; only when the content type field identifies the current data frame as an application layer data protocol, the payload length field value immediately following the content type field is extracted, normalized, and stored in the TLS record layer length sequence, and control data frames of handshake negotiation protocol and alarm protocol type are filtered and discarded; the time-series feature preprocessing module reads the hardware receive timestamp of the data frame, calculates the first-order difference between the timestamp of the current data frame and the timestamp of the previous data frame to obtain the original interval data, and performs natural logarithmic transformation and Z-Score normalization on the original interval data to eliminate the influence of long-tail distribution and generate the message arrival time interval sequence.

[0045] In this embodiment, the edge-aware agent is constructed as an independently running software container or background process, deployed in a bypass monitoring mode on the side-channel observation interface of the data service gateway's network interface card. To ensure zero interference with the gateway's core services, the agent accesses the socket buffer in the kernel space through read-only mapping, without interfering with any forwarding or modification path of service data. This architectural design allows the agent to still mirror and acquire the original traffic data in real time and without loss under high-concurrency traffic, achieving physical isolation between security monitoring and service operation.

[0046] The traffic capture probe module uses zero-copy technology to directly read the binary data stream from the network interface card buffer. Internally, the module runs a state machine that continuously scans the data stream to locate the record layer header structure of the Transport Layer Security (TLS) protocol. Once the header is locked, the module first reads the content type field at the beginning, with pre-defined protocol whitelist filtering logic: if this field indicates that the current data frame is a control type such as handshake negotiation, alarm, or encryption specification change, the module determines it as communication noise and discards it directly to prevent interference with subsequent model analysis. Only when this field indicates an application layer data protocol will the module extract the value of the following payload length field. To adapt to the input requirements of the neural network, the module divides the read byte length value by the maximum record layer length allowed by the protocol or the current network's maximum transmission unit value, mapping it to a floating-point number between 0 and 1, and storing it in the TLS record layer length sequence.

[0047] The timing feature preprocessing module focuses on extracting the physical time fingerprint of traffic. To eliminate jitter errors caused by operating system scheduling, the module directly reads the receive timestamp generated at the network card hardware level for data frames. The module maintains the timestamp record of the previous data frame in memory. By subtracting the timestamp of the previous frame from the timestamp of the current frame, the first-order difference value of the packet arrival time interval is calculated. Considering that network traffic usually has a significant long-tail distribution characteristic, that is, most packet intervals are extremely short, but there are occasional extremely long silent intervals, directly using the original difference value will result in a large numerical range. Therefore, the module performs a natural logarithmic transformation on the original interval data to smooth the data distribution through the compression characteristics of the logarithmic function. Subsequently, the module performs standardization processing on the logarithmically transformed data based on the mean and standard deviation of the sliding window statistics, mapping the data to values ​​that conform to a standard normal distribution, and generating the final packet arrival time interval sequence.

[0048] In practical engineering applications, robust processing logic is designed for specific extreme conditions. First, to address the cold start problem in the early stages of session establishment, when the first data packet arrives and there is no preceding timestamp, the first-order difference value is set to zero or the average value of the current window by default to ensure sequence alignment. Second, to address excessively long silences (such as excessively large heartbeat intervals), a preset threshold is used to truncate the interval value before performing logarithmic transformation to prevent abnormal extreme values ​​from damaging statistical characteristics. Third, regarding out-of-order and retransmitted data packets in the TCP protocol, since this embodiment focuses on the real characteristics of the physical side channel, it strictly records data packets according to the actual physical order in which they arrive at the network card, without logically rearranging out-of-order packets or removing retransmitted packets, thereby ensuring that the model can capture the real physical fluctuations caused by network congestion or attacks. Finally, if a sudden surge in traffic causes a buffer overflow, the module uses a circular buffer mechanism to prioritize overwriting the oldest historical data to ensure that the latest real-time traffic characteristics are always processed.

[0049] Numerical corrections for cold starts and extended periods of silence prevent initial noise and outliers from disrupting statistical patterns and ensure sequence normalization. A ring buffer mechanism addresses sudden traffic surges, ensuring that the latest real-time features are always prioritized even with limited resources, thus improving the model's detection reliability in complex network environments.

[0050] By constructing an independently operating edge-aware agent and employing a bypass side-channel observation mechanism, lossless capture of encrypted traffic is achieved without intruding on the gateway forwarding path or sacrificing service performance. Combined with a precise filtering strategy for application-layer protocols, the system effectively eliminates protocol noise such as handshakes and alarms, improving the signal-to-noise ratio. Furthermore, by performing logarithmic transformation and standardization preprocessing on the time interval sequence, this application successfully eliminates the influence of the long-tail distribution of network traffic data, solving the problem of neural network gradient instability caused by excessively large numerical spans, and providing a stable data foundation for accurately capturing subtle physical side-channel characteristics in encrypted sessions.

[0051] Further, the process of generating the spatiotemporal topology embedding vector includes: the edge-aware agent establishing a temporal-numerical two-dimensional coordinate system, mapping the message arrival time interval sequence to a set of nodes in the coordinate system, where the horizontal axis is the message temporal index and the vertical axis is the time interval value; for any two target nodes in the node set, calculating the linear slope of the line connecting them; if the vertical axis values ​​of all intermediate nodes are lower than the linear interpolation height of the two target nodes at their corresponding temporal positions, then it is determined that the two target nodes have a visible connection relationship, and the corresponding element in the temporal topology adjacency matrix is ​​set to a non-zero connection. The process involves: identifying and constructing a graph attention network with multiple attention heads using an edge-aware agent; mapping the TLS record layer length sequence to high-dimensional node feature vectors using a linear projection matrix; calculating the dot product correlation coefficient between adjacent node feature vectors and using a temporal topological adjacency matrix as a structural mask to set the correlation coefficient of non-connected nodes to negative infinity; normalizing the correlation coefficients after masking to generate an attention weight distribution and weighting and aggregating the node feature vectors in the neighborhood; and concatenating the output vectors of multiple attention heads and generating a spatiotemporal topological embedding vector through a fully connected layer.

[0052] In this embodiment, the edge-aware agent first establishes a virtual time-to-numerical two-dimensional coordinate system in memory, mapping the preprocessed message arrival time interval sequence to a discrete set of points in the coordinate system. The horizontal axis corresponds to the message's time-series index, and the vertical axis corresponds to the normalized time interval value. A linear visibility graph algorithm is used to construct the topological structure of the time-series data. For any two non-adjacent target nodes in the sequence, the system performs geometric visibility detection: calculating the linear slope of the line connecting these two target nodes in the coordinate system, and traversing all intermediate nodes located between the time-series indices of these two nodes. Calculate the linear interpolation height of each intermediate node at its corresponding temporal position (i.e., the theoretical height of the connection at that position); if the actual vertical axis values ​​of all intermediate nodes are strictly lower than the linear interpolation height, it is determined that there is a line-of-sight path between the two target nodes, and they have a visible connection relationship. Based on this determination, the row and column elements corresponding to these two nodes are set to non-zero connection identifiers (e.g., set to 1) in the temporal topology adjacency matrix, otherwise they are set to 0. Through this process, the original one-dimensional time series is transformed into a graph structure with complex network characteristics, in which sudden oscillations in traffic are often transformed into hub nodes with extremely high degrees in the graph.

[0053] After constructing the topology, the edge-aware agent initializes a graph attention network containing multiple independent attention heads. First, for the TLS record layer length sequence as a node attribute, a learnable linear projection matrix is ​​used to perform dimensional expansion, mapping the scalar length value to a high-dimensional node feature vector containing rich semantic information. Then, attention coefficients are calculated. For each central node in the graph, the dot product correlation coefficient between its feature vector and the feature vectors of all neighboring nodes is calculated. To strictly limit attention to the visible range defined by the topology, a temporal topological adjacency matrix is ​​introduced as a structural mask. Specifically, the adjacency matrix is ​​checked, and for node pairs that are not topologically connected, their correlation coefficients are forcibly modified to negative infinity. This operation ensures that the weight probability of non-connected nodes approaches zero in the subsequent Softmax normalization step, thereby blocking the propagation of invalid information.

[0054] The Softmax function is used to normalize the correlation coefficients after masking, generating an attention weight distribution for each neighboring node. This weight reflects which adjacent packet length features are more indicative of the current packet under the current traffic topology. Based on this weight distribution, the system performs a weighted summation of the feature vectors of all nodes in the neighborhood to generate a single attention head context vector.

[0055] To capture traffic patterns in different subspaces (e.g., one set of attention heads focuses on dense bursts of short messages, while another set focuses on periodic heartbeats of long messages), multiple attention heads are run in parallel. Finally, the context vectors output by the multiple attention heads are concatenated along the channel dimension, and the concatenated composite vector is input into a fully connected layer for linear fusion and dimensionality reduction. The final output is a spatiotemporal topology embedding vector that contains both temporal topology structure and load length characteristics.

[0056] The process of generating the spatiotemporal topology embedding vector also includes optimization steps for computational complexity and extreme conditions: the edge-aware agent uses a sliding time window of preset length to segment and truncate the message arrival time interval sequence; it only performs the determination of visibility connection relationships and calculation of connection slopes for the set of nodes located within the current sliding time window, limiting the construction range of the temporal topology adjacency matrix to a local window, so as to control the computational complexity within the linear growth range; for the generated temporal topology adjacency matrix, a compressed sparse row format is used for memory mapping; the matrix elements are traversed, and only non-zero connection identifiers and their corresponding column indices and row offsets are extracted for physical storage, discarding the storage space occupied by zero-value elements; before performing weighted aggregation of the graph attention network, a self-loop injection operation is performed on the temporal topology adjacency matrix, forcibly setting the main diagonal elements of the matrix to non-zero connection identifiers.

[0057] When processing encrypted sessions with long durations or extremely high data throughput, the length of the message arrival time interval sequence accumulates continuously. If visibility geometry determination is performed directly on the entire sequence, the computational load will increase exponentially with the number of messages, causing the gateway to be unable to meet real-time processing requirements. To address this, before constructing the topology, the edge-aware agent first segments the sequence using a pre-defined sliding time window. It then establishes a time-to-numerical coordinate system only within the local area covered by the current sliding window and restricts the determination of visibility connectivity relationships to within the window. Specifically, for any two nodes within the window, it only determines whether the connection between them is blocked by an intermediate node within the same window, without considering historical data outside the window. This strategy forces the time complexity of the graph construction algorithm to grow linearly, ensuring that even when processing long connections and high traffic, the inference latency remains at the millisecond level.

[0058] Temporal topological adjacency matrices generated by linear visibility criteria typically exhibit significant sparsity, meaning that most node pairs do not have direct visible connections. Storing this matrix using a conventional two-dimensional array would consume a large amount of memory, resulting in a significant waste of computational resources by storing zero values ​​representing unconnected nodes. This embodiment employs a compressed sparse row format for physical memory mapping of the adjacency matrix. During storage, the logical adjacency matrix is ​​traversed, discarding all zero-value elements and extracting only non-zero connection identifiers and their corresponding column indices for sequential storage. Simultaneously, a row offset array is maintained to record the starting position of each row's non-zero element in the storage array. This storage method reduces memory usage from being proportional to the square of the number of nodes to being proportional to the actual number of connected edges, significantly reducing the memory overhead of gateway devices and preventing memory overflow failures during traffic surges.

[0059] In extreme conditions where the network is idle or extremely stable, messages may arrive at almost fixed time intervals, forming a flat sequence with completely consistent values. According to the visibility geometry criterion, in such sequences, any connection between non-adjacent nodes will be blocked by intermediate nodes, resulting in a large number of isolated nodes without external connections in the generated topology graph. If such a matrix is ​​directly input into a graph attention network, the isolated nodes will lose their own TLS length features in the weighted calculation because they cannot aggregate neighborhood information, causing the model output to fail. To solve this problem, a self-loop injection operation is performed on the temporal topology adjacency matrix before performing the weighted aggregation operation of the graph attention network. This forces all elements on the main diagonal of the matrix to be marked as non-zero connected. This operation is physically equivalent to ensuring that each node always maintains a connection with itself. Through this mechanism, even in the extreme case where there are no external visible neighbors, the attention mechanism can still allocate attention weights to the node itself, thereby completely preserving and transmitting the original feature information of the node, ensuring the numerical stability and functional availability of the model under various network conditions.

[0060] The multi-head attention network assigns different semantic responsibilities to each attention head to capture different types of periodic features in encrypted traffic: The first group of heads (periodicity detection heads): Each attention head in this group learns a specific query vector and key-value matrix to detect periodic patterns of a specific period length. For example, head 1 focuses on detecting quasi-periodic signals with a period of 5, head 2 focuses on signals with a period of 10, and head 3 focuses on signals with a period of 20. During inference, if a periodicity matching the period of a certain head exists in the current traffic sequence, the attention weight distribution generated by that head will be concentrated at multiples of the period. The second group of heads (burst feature detection heads): This group is used to capture sudden changes in message intervals (i.e., burst features) by calculating the higher-order derivatives of the feature vectors of adjacent nodes. The third group of heads (anomaly detection heads): This group learns the distribution characteristics of background noise and anomalous intervals to identify outliers that deviate from statistical patterns. The output vectors of each attention head are concatenated and then fused and dimensionality reduced through a fully connected layer. During the fusion process, the fully connected layer automatically learns the collaborative relationship between different heads. For example, "if the periodic detection head and the burst feature detection head are activated at the same time, it may indicate periodic mimicry traffic". This forms a unified encoding of periodic fingerprints in the high-dimensional representation space. In this embodiment, the graph attention network contains 9 attention heads, of which 3 are periodic detection heads, 3 are burst feature detection heads, and 3 are anomaly detection heads.

[0061] By introducing a sliding time window truncation mechanism, the time complexity of visibility graph construction is strictly constrained within a linear range, ensuring the real-time inference capability of the gateway device when handling long-term encrypted sessions. Combined with a compressed sparse row memory mapping strategy, the system significantly reduces the redundant storage space of the sparse adjacency matrix, effectively mitigating the risk of memory overflow in scenarios with sudden traffic surges. Furthermore, by utilizing self-loop injection operations, the problem of graph attention feature aggregation failure caused by missing topological connections in extreme conditions such as flat flows is solved, ensuring the numerical stability of the model and the integrity of feature propagation in various complex network environments.

[0062] By establishing a time-numerical coordinate system and applying linear visibility geometry criteria, this application transforms a one-dimensional message arrival time interval sequence into a time-series topological adjacency matrix containing deep dependencies. This effectively reveals the nonlinear physical side-channel characteristics in the time-series fluctuations of encrypted traffic. Combining graph attention networks and structural masking mechanisms, it not only achieves deep integration of TLS record layer length characteristics and time-series topological structures, but also utilizes multi-head attention to dynamically focus on key visible nodes, accurately capturing weak periodic fingerprints in the traffic. This method significantly enhances the model's ability to express features of automated mimicry traffic or low-frequency periodic transmission behavior without the need for decryption, improving the accuracy and robustness of covert side-channel feature detection.

[0063] Furthermore, the edge-aware agent also intercepts the gateway's traffic throughput differential sequence, token bucket remaining capacity, and system load level; the strategy model includes:

[0064] The topology encoder receives the spatiotemporal topology embedding vector as input, uses a residual fully connected network for nonlinear mapping processing, extracts the high-dimensional structured dependencies hidden in the encrypted traffic, and obtains a static topology feature vector.

[0065] The differential trend encoder receives the traffic throughput difference sequence between the current time window and the previous time window of the gateway as input, and uses a one-dimensional convolutional neural network to extract time-series features and generate a dynamic trend feature vector.

[0066] The global state encoder receives the current token bucket remaining capacity and system load level of the gateway as input, and processes them using a preset state embedding layer and a multilayer perceptron to map the resource constraint state of the physical layer into a global constraint feature vector.

[0067] The multimodal gating fusion unit performs a channel-dimensional concatenation operation on the dynamic trend feature vector and the global constraint feature vector to generate an environmental context vector; inputs the environmental context vector into the attention generation network, and uses the Sigmoid activation function to generate a feature recalibration weight vector with a value range of 0 to 1; performs an element-wise Hadamard product operation on the feature recalibration weight vector and the static topological feature vector to output a global content-aware state vector;

[0068] The distribution parameter mapping layer receives the global content-aware state vector and constructs a parallel mean generation linear layer and a standard deviation generation linear layer. The mean generation linear layer maps the input vector to an action mean vector. The standard deviation generation linear layer maps the input vector to an action logarithmic standard deviation vector and performs a natural exponential transformation. The action mean vector and the transformed standard deviation vector are combined to output a multidimensional Gaussian distribution parameter.

[0069] Specifically, in this embodiment, in addition to intercepting the TLS length and time interval of the encrypted session, the edge-aware agent also collects three types of auxiliary data in parallel through the gateway's monitoring interface:

[0070] Static topological features: i.e., high-dimensional spatiotemporal topological embedding vectors output by the graph attention network in the preceding steps;

[0071] Dynamic trend characteristics: The real-time throughput of the gateway is sampled at fixed time windows (e.g., 1 second), and the difference sequence between the throughput of the current window and the previous window is calculated. This sequence reflects the sudden fluctuation trend of traffic.

[0072] Global state characteristics: Real-time reading of the current token counter value (remaining capacity) in the token bucket algorithm and the average CPU load of the operating system, and normalization of these two physical quantities to the [0, 1] interval.

[0073] The topology encoder is mainly responsible for further nonlinear refinement of the deep graph features output by GAT. To prevent the gradient vanishing problem in deep networks, this embodiment adopts a residual fully connected network structure. Specifically, the input spatiotemporal topology embedding vector first undergoes dimensionality transformation through a fully connected layer and is nonlinearly activated by ReLU or Swish activation functions. Subsequently, it is restored to the original dimension through a second fully connected layer, and the output of this layer is added element-wise to the original input vector to form a skip connection. This structure ensures that the model can learn new nonlinear transformation features while retaining the original graph topology information, outputting a static topology feature vector containing high-dimensional structured dependencies.

[0074] The specific implementation of the topology encoder includes: the input spatiotemporal topology embedding vector first passes through a first fully connected layer, expanding its dimension from the embedding vector dimension d_e to d_e×2 (e.g., if the spatiotemporal topology embedding vector dimension is 128, it is expanded to 256) to enhance feature representation capability; nonlinear activation is performed using the ReLU activation function, and when the network depth exceeds 4 layers, the Swish activation function is used to improve gradient flow characteristics; subsequently, a second fully connected layer restores the dimension to the original d_e dimension. Layer Normalization is added between the two fully connected layers to stabilize training, and dimension alignment is performed for skip connections. If the output dimension of the second fully connected layer is inconsistent with the original input dimension, a linear projection matrix is ​​used to adjust the dimension of the input vector, followed by element-wise addition; finally, a final nonlinear transformation is performed through an output fully connected layer to generate a static topology feature vector containing high-dimensional structured dependencies.

[0075] The differential trend encoder aims to capture the instantaneous change patterns of traffic waveforms. The traffic throughput differential sequence is constructed as a single-channel time-series input and fed into a one-dimensional convolutional neural network. This network contains multiple sets of one-dimensional convolutional kernels of different sizes, used to slide and scan along the time dimension to extract rising edges, falling edges, or oscillation patterns of the traffic. The convolutional output is then subjected to dimensionality reduction through a max-pooling layer to extract the most significant feature responses. Finally, it is flattened through a flattening layer to generate a dynamic trend feature vector reflecting the current dynamic change pattern of the traffic.

[0076] The global state encoder processes low-dimensional physical constraint data. First, it uses a pre-set state embedding layer to map the normalized token bucket remaining capacity and system load level to a high-dimensional vector space. Specifically, it uses a learnable projection matrix to expand these two scalars into dense vectors. Then, it uses a multilayer perceptron (MLP) to perform feature interaction, transforming the physical layer's resource red lines (such as token exhaustion or CPU overload) into semantic constraints that the model can understand, and generating a global constraint feature vector.

[0077] To achieve intelligent decision-making that dynamically adjusts the focus based on environmental conditions, a multimodal gating mechanism is introduced. First, the dynamic trend feature vector representing traffic trends and the global constraint feature vector representing resource constraints are concatenated along the channel dimension to generate a composite vector containing current environmental context information. Subsequently, this composite vector is input into the attention generation network (usually a lightweight fully connected layer) and processed by the Sigmoid activation function. The Sigmoid function strictly compresses the output value between 0 and 1, generating a feature recalibration weight vector. Finally, the system performs a crucial gating operation: the weight vector is subjected to an element-wise Hadamard product operation with the static topology feature vector. This step is equivalent to a soft switch: when the environmental context indicates that the network is extremely congested or the traffic features are chaotic, the gating mechanism automatically generates weights close to 0, suppressing complex topology features and forcing the model to adopt a conservative strategy; otherwise, it preserves topology details, and the result is the global content-aware state vector.

[0078] The final output layer of the policy model is designed as a dual-path parallel linear layer to construct the stochastic policy distribution required for reinforcement learning.

[0079] Mean generation branch: Receives the global content-aware state vector and directly outputs the action mean vector through linear projection. This vector represents the token adjustment coefficient that the model considers to be optimal at the moment.

[0080] Standard deviation generation branch: Receives the same state vector, outputs the logarithmic standard deviation vector of the action through linear projection. In order to ensure the strict non-negativity of the standard deviation in mathematics (the standard deviation cannot be negative), a natural exponential transformation is performed on the logarithmic standard deviation to obtain the actual action standard deviation vector. Finally, a multidimensional Gaussian distribution is constructed using the above mean vector and standard deviation vector. The flow agent will sample from this distribution to generate exploratory final control commands.

[0081] The specific judgment logic of the linear visibility geometric criterion is as follows: First, in the established temporal-numerical two-dimensional coordinate system, any two non-overlapping nodes are selected as target nodes, with the one earlier in the temporal sequence being the starting node and the one later in the temporal sequence being the ending node. The ratio of the numerical difference between these two target nodes on the vertical axis to the temporal index difference on the horizontal axis is calculated. This ratio is the linear slope of the straight line connecting the two target nodes. Second, for each intermediate node located between the starting node and the ending node, its corresponding linear interpolation height is calculated. The linear interpolation height is defined as: the vertical axis value of the starting node, plus the product of the linear slope and the temporal distance (i.e., the temporal index difference) of the intermediate node relative to the starting node. This height represents the theoretical vertical axis value that the intermediate node should have if it is located exactly on the line connecting the starting node and the ending node. Third, a line-of-sight accessibility test is performed: the actual vertical axis value of each intermediate node is compared with the corresponding linear interpolation height calculated above. Only when the actual vertical axis values ​​of all intermediate nodes are strictly less than their corresponding linear interpolation heights is it determined that the starting and ending nodes are not obstructed and have a visible connection. In this case, the corresponding connection flag is set to valid in the temporal topology adjacency matrix. In particular, to handle computational accuracy errors and boundary conditions, the system adopts the following supplementary rules: First, for two nodes that are temporally adjacent (i.e., there are no other nodes in between), no interpolation calculation is required, and they are directly determined to have a visible connection. Second, when comparing the actual value of an intermediate node with the linear interpolation height, a very small preset numerical tolerance range is introduced. If the difference between the two is within this tolerance range, they are considered equal and treated as if the line of sight is obstructed, in order to avoid logical misjudgments caused by computer floating-point arithmetic.

[0082] By constructing a strategy model based on multi-view feature fusion, the deep representation capability of residual fully connected networks for static topology and the dynamic capture capability of one-dimensional convolutional networks for traffic trends are combined. Using multimodal gating fusion units, feature recalibration weights are generated according to global resource constraints and traffic oscillation context, and dynamic filtering and noise suppression are performed on topological features, which solves the problem of insufficient robustness of single features under complex working conditions. In addition, the dual-path decoupled output and standard deviation exponential transformation mechanism are adopted to ensure the mathematical validity of the action probability distribution and improve the convergence stability of the strategy gradient algorithm in continuous action space exploration.

[0083] Furthermore, the process of outputting the response traffic flow and recording the gateway processing latency metric introduced by asynchronous latency includes: the traffic agent employs a reparameterization technique to sample random noise from a standard normal distribution and calculates the token adjustment coefficient by combining it with the distribution parameters output by the policy model; it reads the gateway's preset baseline token delivery rate, multiplies the baseline token delivery rate with the token adjustment coefficient to generate the dynamic filling rate within the current time window; it obtains the payload bytes of the data packet to be sent and queries the remaining token count in the current token bucket; when the payload bytes are greater than the remaining token count, it calculates the difference between the two. The difference is divided by the dynamic filling rate to calculate the waiting time required for the data packet to acquire a token equal to the number of bytes in the data packet payload; the current system kernel time is read and added to the waiting time to determine the target release timestamp of the data packet; a delayed task object containing a data packet reference and a callback function is constructed, and the task object is registered to the corresponding time slot of the hierarchical time wheel scheduler according to the target release timestamp; when the system time advances to the target release timestamp, the scheduler triggers the callback function, writes the data packet into the gateway sending queue, and accumulates the waiting time into the statistical buffer of the gateway processing latency index.

[0084] To enable effective gradient backpropagation of reinforcement learning policies in the continuous action space, the traffic agent employs a reparameterization technique to generate control commands. Specifically, it first randomly samples a noise scalar from a pre-defined standard normal distribution (mean 0, variance 1). Then, it reads the action mean and action standard deviation output by the policy model, multiplies the noise scalar by the action standard deviation, and adds the action mean to generate a token adjustment coefficient with stochastic exploration. This coefficient directly affects the gateway's flow control core. It reads the preset baseline token delivery rate (e.g., 1MB tokens per second) from the gateway configuration file, multiplies the baseline rate by the aforementioned token adjustment coefficient, and calculates the dynamic filling rate within the current time window. In this way, the agent can dynamically accelerate or slow down token generation based on the current traffic topology characteristics, thereby controlling the speed of traffic flow.

[0085] Traditional token bucket algorithms are typically used for rate limiting, where packets are dropped when there are insufficient tokens. To achieve traffic shaping required for covert channel defense, this embodiment employs a reverse token bucket logic. When there are insufficient tokens, packets are queued and intercepted. The number of bytes in the payload (i.e., the number of tokens required for the packet) is obtained, and the remaining tokens in the current token bucket counter are checked. If the number of bytes in the payload is less than or equal to the number of remaining tokens, the packet is allowed to pass directly. If the number of bytes in the payload is greater than the number of remaining tokens, the difference between the two is calculated, i.e., the token deficit. Then, the time required to fill the deficit is calculated using a physical formula: the token deficit is divided by the current dynamic filling rate to obtain the precise duration that the packet must wait. This step transforms the AI's decision (rate) into a constraint (time) in the physical world.

[0086] To efficiently manage delayed tasks of massive concurrent data packets with microsecond-level precision without blocking the gateway's core I / O thread, this embodiment employs a hierarchical time wheel scheduler. First, the current system kernel time is read and added to the calculated waiting time to determine the absolute target release timestamp of the data packet. Then, a delayed task object is constructed, which encapsulates the memory reference of the data packet and the sending callback function. Based on the difference between the target release timestamp and the current time, the task object is registered to the corresponding level and slot of the time wheel. The hierarchical time wheel consists of multiple circular arrays with different time granularities (such as microsecond, millisecond, and second levels). Tasks are managed efficiently by rotating pointers, avoiding the O(logN) complexity overhead of traditional priority queues during insertion and deletion, and ensuring that the time complexity of the scheduling operation is close to O(1).

[0087] As the system clock advances, when the time wheel pointer points to the slot where the task object is located, the scheduler automatically triggers the callback function, pushing the corresponding data packet back into the gateway's network card sending queue or TCP protocol stack to complete the physical transmission. At the same time, at the moment the callback function is executed, the actual waiting time experienced by the data packet (i.e., the artificial delay introduced by the shaping) is added to the statistical buffer of the gateway's processing latency indicator. This indicator will serve as a key negative feedback signal for subsequent reflection on the agent's evaluation of the strategy's merits (i.e., whether it has excessively affected business performance), ensuring a balance between security defense and business experience.

[0088] The callback function in the delayed task object is a predefined executable code segment whose logic is as follows: It uses a spinlock to protect critical access to the shared state of the traffic agent, ensuring atomic operations; it extracts the memory reference of the data packet and the pointer to the network card's sending queue from the task object; it copies the memory content of the data packet to the tail of the sending queue; it executes a memory barrier instruction to ensure visibility; it updates the tail pointer of the sending queue, enabling the network card driver to detect the arrival of new data packets; it reads the system clock to obtain the current precise time, compares it with the target release timestamp, calculates the deviation between the actual delay and the planned delay (usually in the microsecond range), and accumulates this deviation into the atomic counter of the gateway's latency metric; it releases the spinlock. In extreme cases, if the current system time has far exceeded the target release timestamp (e.g., a difference of more than 100 microseconds), the callback function executes immediate sending without waiting and records a delay expiration alarm event.

[0089] By employing a reparameterized sampling mechanism, the policy gradient algorithm achieves efficient exploration in the continuous action space, generating token adjustment instructions that dynamically adapt to the current network environment. Combined with the inverse token bucket algorithm, the abstract flow control rate is transformed into precise packet waiting time, reconstructing traffic timing characteristics without packet loss and effectively interfering with side-channel analysis. Furthermore, a hierarchical time wheel scheduler is used to execute microsecond-level asynchronous delay operations, reducing the scheduling complexity of massive delay tasks to a constant level, avoiding thread blocking in high-concurrency scenarios, and ensuring that the data service gateway maintains high throughput performance when executing complex traffic shaping strategies.

[0090] Further, the backpropagation update of the policy model parameters using the stochastic gradient ascent algorithm includes: adjusting the policy model parameters to maximize the substitution objective function using the stochastic gradient ascent algorithm includes: the red team agent extracts the statistical feature vector of the response traffic flow and inputs it into a preset random forest classifier; the classifier contains multiple independently trained decision trees, each decision tree outputs a business category prediction result; the output results of all decision trees are statistically analyzed, and the voting percentage of the correct business label is calculated as the recognition confidence; the difference between 1 and the recognition confidence is calculated, and the difference is determined as the traffic confusion escape rate; the reflective agent uses a weighted linear combination method and a weighted calculation method to map the normalized traffic confusion escape rate as a positive reward term and the normalized gateway processing latency indicator as a negative penalty term, and calculates the difference between the positive reward term and the negative penalty term, and determines the difference as the instant reward value at the current moment; using a preset... The value network predicts the state value scalars of the current state and the state at the next time step; it combines the immediate reward value with the state value scalar to calculate the temporal difference error; it introduces a discount factor and a smoothing parameter, and performs an exponentially weighted moving average calculation on multiple temporal difference errors within a future time window to obtain the generalized dominance function estimate; it reads the parameters of the policy model before and after the update, calculates the probability of the new policy and the probability of the old policy to perform the same action, and calculates the ratio between the two; it calculates the product of the ratio and the generalized dominance function estimate to obtain the untruncated target term; simultaneously, it restricts the ratio to a preset floating range centered at 1, calculates the product of the restricted ratio and the generalized dominance function estimate to obtain the truncated target term; it selects the minimum value between the untruncated target term and the truncated target term as the alternative target function, uses the stochastic gradient ascent algorithm to calculate the gradient of the alternative target function with respect to the policy model parameters, and performs backpropagation to update the policy model parameters.

[0091] Specifically, in this embodiment, the red team agent acts as an independent evaluation module to detect the response traffic flow output by the traffic agent in real time. First, it extracts statistical features from the response traffic flow within the current time window to generate a statistical feature vector containing statistics on packet payload length (such as mean and variance) and the distribution characteristics of message arrival time intervals. Second, it inputs the statistical feature vector into a preset random forest classifier, which contains multiple independently trained decision trees. Each decision tree independently predicts the service category to which the current traffic belongs based on the input feature vector and outputs a prediction label. Finally, it calculates the prediction results of all decision trees and the proportion of decision trees whose prediction results match the actual service labels to the total number of trees. This proportion is determined as the recognition confidence. The difference between the value 1 and the recognition confidence is calculated to obtain the traffic confusion escape rate. This value directly reflects the confusion effect of the traffic shaping strategy against the random forest model.

[0092] The specific logic for the reflective agent to calculate the immediate reward value at the current moment is as follows: First, two scalar weight coefficients are preset to balance security and performance, defined as security benefit weight and performance cost weight respectively (for example, in scenarios where security is emphasized, they can be set to 0.7 and 0.3 respectively). Second, the agent multiplies the received traffic obfuscation escape rate (value range 0 to 1) with the security benefit weight to obtain a positive reward component. At the same time, it multiplies the monitored gateway processing latency indicator (unit: milliseconds) with the performance cost weight to obtain a cost component representing the penalty. Finally, it performs an algebraic operation, subtracting the cost component from the positive reward component, and the final scalar value obtained is the immediate reward value required for reinforcement learning. Through this calculation method, it is clear that the agent must try to reduce the impact on business latency while improving the obfuscation success rate. Any excessive increase in latency will lead to a significant decrease in the total reward value due to the increased minuend.

[0093] To ensure that the reflective agent can reasonably evaluate the merits of its current actions, an adaptive reward calculation mechanism based on normalization and dynamic weight awareness was constructed. The specific execution process of this mechanism includes three consecutive steps: indicator normalization, dynamic weight determination, and final reward aggregation.

[0094] The first step is to perform normalization preprocessing for heterogeneous metrics. Since the traffic obfuscation escape rate is a probability value between 0 and 1, while the gateway processing latency metric is a physical quantity measured in milliseconds, to eliminate the difference in dimensions, a maximum tolerable latency threshold is first set (e.g., 500 milliseconds, representing the minimum acceptable performance level for the business). The actual gateway processing latency value monitored in real time is then compared with this maximum tolerable latency threshold, mapping the millisecond-level physical time to a latency penalty coefficient between 0 and 1. Specifically, if the actual latency exceeds this threshold, the coefficient is forcibly truncated and locked to 1 to prevent distortion of the reward function due to numerical overflow.

[0095] The second step involves dynamically determining weights based on two-dimensional perception. First, an initial weight template is loaded according to the current business scenario. For example, in a high-level protection mode, a higher security benefit weight and a lower performance cost weight are initialized. Then, during operation, the optimization module uses two-dimensional perception logic to fine-tune these weights in real time: First, based on threat situation assessment, if the traffic obfuscation escape rate reported by the red team agent remains above the warning line for multiple consecutive time windows, indicating insufficient defense, the security benefit weight will be automatically increased gradually in fixed steps, forcing the model to adopt a more aggressive strategy. Second, based on system load assessment, if the gateway's CPU or memory resource utilization exceeds the safety threshold, indicating an overload risk, the system will automatically decrease the security benefit weight and correspondingly increase the performance cost weight to guide the model towards convergence towards lower resource consumption. At all times, the system ensures that the algebraic sum of the security benefit weight and the performance cost weight remains constant.

[0096] The third step is to perform weighted difference aggregation, which multiplies the original traffic obfuscation escape rate with the currently determined security benefit weight to obtain a positive benefit term; and multiplies the normalized latency penalty coefficient with the currently determined performance cost weight to obtain a negative penalty term. Finally, the difference between the positive benefit term and the negative penalty term is calculated, and this difference is determined as the instantaneous reward value at the current moment, thereby realizing a dynamic game and balance between security effect and performance loss within a unified numerical space.

[0097] To assess the advantage of the current action relative to the average policy, the reflexive agent evaluates the environmental state using a pre-defined value network. First, the current state vector and the state vector of the next time step are input into the value network to obtain the corresponding predicted state value scalar values. Second, the temporal difference error is calculated, which is equal to the immediate reward value of the current time step plus the discounted state value scalar value of the next time step, and then subtracted from the state value scalar value of the current time step. Finally, to balance the estimation bias and variance, a discount factor and a smoothing parameter are introduced, and an exponentially weighted moving average is calculated for multiple temporal difference errors generated continuously within a preset future time window. The accumulated result is determined as the generalized advantage function estimate.

[0098] The proximal policy optimization algorithm is used to update the parameters of the policy model. First, the old parameters and the new parameters to be optimized in the policy model are read, and the probability values ​​of performing the same action in the current state are calculated. The ratio of the new policy probability to the old policy probability, i.e., the action probability ratio, is calculated. Second, two objective terms are constructed for comparison: the first term is the untruncated objective term, which is obtained by directly multiplying the action probability ratio by the generalized dominance function estimate; the second term is the truncated objective term, which is obtained by forcibly restricting the action probability ratio to a preset floating range centered at the value 1, and then multiplying the restricted ratio by the generalized dominance function estimate. Third, the minimum value between the untruncated and truncated objective terms is selected as the final alternative objective function. Finally, the gradient of the alternative objective function with respect to the policy model parameters is calculated using the stochastic gradient ascent algorithm, and backpropagation is performed to update the weight matrix and bias vector in the policy model, completing one policy iteration.

[0099] The specific calculation logic for the generalized advantage function valuation is as follows: First, the key hyperparameters are clearly defined: the discount factor is set to 0.99, which measures the importance of future long-term rewards to the current decision; the closer the value is to 1, the more emphasis is placed on long-term returns. The smoothing parameter is set to 0.95, which is used to balance bias and variance. The so-called future time window is specifically set to 5 seconds in this embodiment. If calculated with the system's control frequency of 20Hz, it covers 100 consecutive sampling steps in the future. Second, the specific weighted calculation is performed: first, the temporal difference error of each time step is calculated, which reflects the deviation between the immediate reward brought by the current action and the expected value. Subsequently, in order to obtain the current generalized advantage function estimate, the system exponentially weights and accumulates all time-series difference errors starting from the current moment and covering the future time window. The specific weighting rule is as follows: the discount factor is multiplied by the smoothing parameter to obtain a mixed decay coefficient less than 1 (approximately 0.94 in this embodiment). For future error terms that are further away from the current moment, their weight is the power of this mixed decay coefficient. This means that the contribution of the prediction error of the more distant future to the current estimate decreases exponentially. Through this weighting method described in words, the return information of multiple future steps is effectively integrated into the current decision evaluation, which not only retains the accuracy of the Monte Carlo method but also introduces the stability of the time-series difference method.

[0100] The statistical feature vector extracted by the red team agent includes: the mean, standard deviation, minimum, maximum, kurtosis, and skewness of the payload length of all application layer data packets within a single time window; the mean, standard deviation, minimum, maximum, autocorrelation coefficient, and Hurst exponent of the interval sequence of message arrival time intervals; the number of data packets per unit time, the coefficient of variation of the data packet arrival rate, and the traffic time-series uncertainty quantized using Shannon entropy; and the main frequency component and its energy proportion of the traffic sequence detected using Fourier transform. These features, totaling 20 dimensions, constitute the statistical feature vector.

[0101] The random forest classifier consists of 100 independently trained decision trees, with a maximum depth of 16 layers per tree to avoid overfitting. The training set of the classifier consists of encrypted traffic samples labeled offline, including three categories: (1) human-interactive traffic (label=0), characterized by high entropy and irregularity in the interval sequence; (2) automated program traffic (label=1), characterized by periodicity or quasi-periodicity in the interval sequence; and (3) background noise / unknown traffic (label=2).

[0102] To cope with the dynamic changes in the network environment, the reflective agent periodically (every 24 hours) incrementally learns the classifier while iterating its strategy. It manually labels traffic samples with low confidence (between 0.6 and 0.8) identified in a historical day, updates the classifier's training set, and retrains the random forest model.

[0103] After completing a policy iteration, the reflective agent synchronously deploys the updated policy model parameters (including weight matrix and bias vector) to the inference engine of the traffic agent in real time, replacing the original old parameters. The traffic agent then uses the optimized new policy model to perform a new round of dynamic traffic shaping operations on the encrypted sessions of the data service gateway that arrive subsequently.

[0104] At the same time, the reflective agent performs a buffer cleanup operation, clearing the historical interaction data generated based on the old strategy in the experience replay buffer and resetting the sampling counter. In this way, it enters the next sampling-evaluation-update closed-loop cycle. Through this continuous self-game and parameter fine-tuning, the multi-agent system can enable the defense strategy to gradually converge to the optimal state of balancing security benefits and business performance in a dynamically changing network adversarial environment, thereby achieving autonomous evolution and adaptive enhancement of security operation capabilities.

[0105] By constructing a closed-loop feedback mechanism based on red team adversarial evaluation and policy gradient optimization, and utilizing a random forest classifier to quantify traffic obfuscation escape rate in real time, an objective concealment evaluation index is provided for the policy model. Combined with a multi-dimensional reward function, the gateway processing latency is effectively constrained while pursuing high security gains, achieving a dynamic balance between security defense and business performance. Furthermore, the generalized advantage function estimation and probability ratio truncation update strategy are adopted to smooth the variance of gradient estimation and limit the magnitude of single updates, effectively preventing network service jitter caused by drastic policy oscillations, and ensuring the numerical stability and continuous evolution capability of the agent in complex adversarial environments.

[0106] The effectiveness of the linear visibility criterion in identifying mimicry traffic is closely related to the obfuscation strategies employed by attackers. This invention primarily targets the following types of mimicry traffic for defense: Attackers transmit data at fixed or near-fixed time intervals, which appear as flat or low-fluctuation straight lines in a time-to-numerical coordinate system; the linear visibility criterion can effectively distinguish this as a sparse topology. Automated programs, when implementing mimicry, often generate time intervals based on a set rule engine; these rules typically exhibit segmented, constant periodicity, fundamentally different from the interaction interval distribution of real users; the linear visibility criterion can identify this through topological sparsity. Attackers intentionally use extremely low transmission frequencies or extremely long transmission periods to evade detection based on traffic characteristics; the time interval sequences of this type of traffic exhibit long-period periodicity, and the linear visibility criterion, with the assistance of multi-head attention, can capture these weak periodic signals. Therefore, the defensive effect of this invention is dynamic and adaptive, and does not provide complete protection against all possible attacks.

[0107] By constructing a security operation system based on multi-agent collaboration, this application achieves accurate perception and adaptive defense of the concealed channel of encrypted sessions without decrypting business content. Utilizing the linear visibility criterion and graph attention mechanism, physical side-channel characteristics such as message timing are transformed into high-dimensional topological representations, effectively identifying low-frequency, slow attacks that are difficult to capture by traditional rules. Combined with policy gradient-driven dynamic traffic shaping technology, this application can reconstruct traffic fingerprints in real time with microsecond-level accuracy. Through adversarial noise filling and asynchronous delay scheduling, it successfully interferes with the attacker's traffic analysis behavior. Furthermore, based on the closed-loop feedback mechanism of red team escape rate assessment, a dynamic balance between security benefits and business performance is established, enabling the continuous evolution and robust operation of the defense strategy in complex network environments.

[0108] Example 2:

[0109] This embodiment applies to a high-frequency data service interface (such as an encrypted file transfer interface) in a city's public data sharing space. The main technical challenge faced by this interface is the presence of a large amount of automated machine traffic. This traffic attempts to bypass gateway policies based on fixed frequency restrictions by simulating the random access intervals of normal users (mimicry). This mimicry traffic uses TLS encryption throughout its transmission and statistically exhibits a long-term, implicit pattern with random masking, making it difficult to identify and isolate using traditional stateless firewalls. To improve data security, a multi-agent collaborative security operation system provided in this application is introduced. Edge-aware agents, traffic agents, red-team agents, and reflective agents are deployed in the data gateway as embedded middleware, responsible for real-time control of the physical side-channel characteristics of the encrypted channel.

[0110] When an external client establishes an encrypted session with the gateway, the edge-aware agent intercepts the encrypted response data stream through the underlying driver. The agent extracts the payload length sequence of the TLS record layer, which reflects the volume fluctuation characteristics of the data content. Simultaneously, the agent records the message arrival time interval sequence. Addressing the technical characteristic that simulated traffic, despite randomization, still retains underlying mechanical periodicity, the agent uses linear visibility geometry criteria to map the time interval sequence into a temporal topological adjacency matrix. In this topology, the mechanical periodicity of automated programs is mapped as a specific regularized chain subgraph, while real human interaction traffic is presented as a scale-free fractal subgraph. The graph attention network uses structural masks to aggregate these micro-topological features, generating a high-dimensional spatiotemporal topological embedding vector, thereby physically distinguishing between machine-generated traffic and human interaction traffic.

[0111] The traffic agent receives the aforementioned embedded vector. The policy model identifies that the current session has high machine-generated characteristics, mapping a low token generation rate adjustment coefficient and a large action standard deviation. The agent uses this coefficient to dynamically reduce the frequency of token distribution to the session by the gateway and triggers the reverse token bucket logic: when a data packet arrives, due to insufficient tokens, the system calculates the microsecond-level waiting time required to fill the deficit. The agent registers the data packet with the hierarchical time wheel scheduler and executes asynchronous delayed transmission. This process introduces nonlinear physical timing jitter at the transport layer, disrupting the original time interval pattern of the traffic. This makes the originally periodic pseudo-traffic discrete and disordered in the time domain, eliminating its inherent side-channel fingerprint.

[0112] The red team agent uses a pre-built traffic fingerprint classifier (built on random forest) to detect the shaped output traffic. Before shaping, the classifier can classify the traffic as automated traffic based on its temporal periodicity. After shaping, due to the introduction of physical temporal jitter, the classifier's recognition confidence drops significantly, misclassifying it as unknown type or background noise, thus generating a high traffic confusion escape rate. The agent reads the average processing latency index introduced by the gateway, calculates the generalized dominance function estimate based on the high escape rate, and uses a probability ratio truncation mechanism to update the agent's strategy model parameters. This continuously strengthens the transmission control strategy that applies dynamic jitter to specific temporal characteristics, achieving effective suppression and feature cleaning of automated mimicry traffic without interrupting the connection.

[0113] 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 safety operation method based on multi-agent collaboration, characterized in that, include: The edge-aware intelligent agent intercepts the TLS record layer length sequence and message arrival time interval sequence of the encrypted session from the data service gateway; Based on the linear visibility geometry criterion, the time interval sequence is mapped to a temporal topological adjacency matrix. The attention mechanism is used to weight the length sequence and the matrix to generate a spatiotemporal topological embedding vector. The embedded vector input strategy model is mapped to multidimensional Gaussian distribution parameters. The distribution parameters are sampled by the traffic agent to generate token adjustment coefficients. The token distribution frequency of the gateway token bucket algorithm is updated using the adjustment coefficients. The target release timestamp of the data packets to be sent is calculated based on the current token status in the bucket. Asynchronous delay is performed on data packets that have not reached the timestamp. The response traffic stream is output, and the gateway processing delay index introduced by the asynchronous delay is recorded. The red team agent uses random forest to perform business fingerprinting on the response traffic flow and generates traffic confusion escape rate; the reflective agent reads latency indicators and escape rate, calculates the generalized advantage function estimate and the action probability ratio of the old and new strategies, constructs the substitution objective function using a preset cutoff threshold, and adjusts the parameters of the strategy model through the stochastic gradient ascent algorithm to maximize the substitution objective function.

2. The safety operation method based on multi-agent collaboration according to claim 1, characterized in that, The process by which the edge-aware agent intercepts data from the data service gateway includes: the edge-aware agent is constructed as an independent running instance containing a traffic capture probe module and a time-series feature preprocessing module, and deployed on the network-side channel observation interface of the data service gateway; the traffic capture probe module reads binary data frames in the gateway network buffer in real time, locates the secure transport layer protocol record layer header structure, and parses the content type field in the header structure; only when the content type field identifies the current data frame as an application layer data protocol, the payload length field value immediately following the content type field is extracted, normalized, and stored in the TLS record layer length sequence, and control data frames of handshake negotiation protocol and alarm protocol type are filtered and discarded; the time-series feature preprocessing module reads the hardware receive timestamp of the data frame, calculates the first-order difference between the timestamp of the current data frame and the timestamp of the previous data frame to obtain the original interval data, and performs natural logarithmic transformation and Z-Score normalization on the original interval data to eliminate the influence of long-tail distribution and generate the message arrival time interval sequence.

3. The safety operation method based on multi-agent collaboration according to claim 1, characterized in that, The process of generating the spatiotemporal topology embedding vector includes: the edge-aware agent establishing a temporal-numerical two-dimensional coordinate system, mapping the message arrival time interval sequence to a set of nodes in the coordinate system, where the horizontal axis is the message time sequence index and the vertical axis is the time interval value; for any two target nodes in the node set, calculating the linear slope of the line connecting the two target nodes; if the vertical axis values ​​of all intermediate nodes are lower than the linear interpolation height of the two target nodes at their corresponding temporal positions, then it is determined that the two target nodes have a visible connection relationship, and the corresponding element in the temporal topology adjacency matrix is ​​set to a non-zero connection. The process involves: identifying and constructing a graph attention network with multiple attention heads using an edge-aware agent; mapping the TLS record layer length sequence to high-dimensional node feature vectors using a linear projection matrix; calculating the dot product correlation coefficient between adjacent node feature vectors and using a temporal topological adjacency matrix as a structural mask to set the correlation coefficient of non-connected nodes to negative infinity; normalizing the correlation coefficients after masking to generate an attention weight distribution and weighting and aggregating the node feature vectors in the neighborhood; and concatenating the output vectors of multiple attention heads and generating a spatiotemporal topological embedding vector through a fully connected layer.

4. The safety operation method based on multi-agent collaboration according to claim 1, characterized in that, The edge-aware agent also intercepts the gateway's traffic throughput differential sequence, token bucket remaining capacity, and system load level; the policy model includes: The topology encoder receives the spatiotemporal topology embedding vector as input, uses a residual fully connected network for nonlinear mapping processing, extracts the high-dimensional structured dependencies hidden in the encrypted traffic, and obtains a static topology feature vector. The differential trend encoder receives the traffic throughput difference sequence between the current time window and the previous time window of the gateway as input, and uses a one-dimensional convolutional neural network to extract time-series features and generate a dynamic trend feature vector. The global state encoder receives the current token bucket remaining capacity and system load level of the gateway as input, and processes them using a preset state embedding layer and a multilayer perceptron to map the resource constraint state of the physical layer into a global constraint feature vector. The multimodal gating fusion unit performs a channel-dimensional concatenation operation on the dynamic trend feature vector and the global constraint feature vector to generate an environmental context vector; inputs the environmental context vector into the attention generation network, and uses the Sigmoid activation function to generate a feature recalibration weight vector with a numerical range of 0 to 1; performs an element-wise Hadamard product operation on the feature recalibration weight vector and the static topological feature vector to output a global content-aware state vector. The distribution parameter mapping layer receives the global content-aware state vector and constructs a parallel mean generation linear layer and a standard deviation generation linear layer. The mean generation linear layer maps the input vector to an action mean vector. The standard deviation generation linear layer maps the input vector to an action logarithmic standard deviation vector and performs a natural exponential transformation. The action mean vector and the transformed standard deviation vector are combined to output a multidimensional Gaussian distribution parameter.

5. A safety operation method based on multi-agent collaboration according to claim 1, characterized in that, The process of outputting the response traffic flow and recording the gateway processing latency metrics introduced by asynchronous delay includes: the traffic agent sampling random noise from a standard normal distribution using reparameterization, and calculating the token adjustment coefficient by combining it with the multidimensional Gaussian distribution parameters output by the policy model; reading the gateway's preset baseline token delivery rate, multiplying the baseline token delivery rate with the token adjustment coefficient to generate the dynamic filling rate within the current time window; obtaining the payload bytes of the data packet to be sent, and querying the remaining tokens in the current token bucket; when the payload bytes are greater than the remaining tokens, calculating the difference between the payload bytes and the remaining tokens. The system calculates the waiting time required to acquire a token equal to the number of bytes in the data packet payload by dividing the difference by the dynamic filling rate. It then reads the current system kernel time and adds it to the waiting time to determine the target release timestamp of the data packet. A delayed task object containing a data packet reference and a callback function is constructed and registered to the corresponding time slot in the hierarchical time wheel scheduler based on the target release timestamp. When the system time advances to the target release timestamp, the scheduler triggers the callback function, writes the data packet to the gateway sending queue, and accumulates the waiting time into the statistical buffer of the gateway's processing latency metrics.

6. A security operation method based on multi-agent collaboration according to claim 1, characterized in that, The step of adjusting the parameters of the strategy model using the stochastic gradient ascent algorithm to maximize the substitution objective function includes: the red team agent extracting statistical feature vectors of the response traffic flow and inputting them into a pre-set random forest classifier; the classifier contains multiple independently trained decision trees, each outputting a business category prediction result; the output results of all decision trees are statistically analyzed, and the voting percentage of the correct business label is calculated as the recognition confidence; the difference between 1 and the recognition confidence is calculated, and the difference is determined as the traffic confusion escape rate; the reflective agent uses a weighted linear combination method and a weighted calculation method to map the normalized traffic confusion escape rate as a positive reward term and the normalized gateway processing latency index as a negative penalty term, and calculates the difference between the positive reward term and the negative penalty term, and determines the difference as the immediate reward value at the current moment; the current state and the state at the next moment are predicted using a pre-set value network. The state value scalar is used; the temporal difference error is calculated by combining the immediate reward value and the state value scalar; a discount factor and smoothing parameter are introduced, and an exponentially weighted moving average is performed on multiple temporal difference errors within a future time window to obtain the generalized dominance function estimate; the parameters of the policy model before and after the update are read, and the probabilities of the new policy and the old policy for performing the same action are calculated respectively, and the ratio between the two is calculated; the product of the ratio and the generalized dominance function estimate is calculated to obtain the untruncated target term; at the same time, the ratio is restricted to a preset floating range centered at 1, and the product of the restricted ratio and the generalized dominance function estimate is calculated to obtain the truncated target term; the minimum value between the untruncated target term and the truncated target term is selected as the alternative target function, and the gradient of the alternative target function with respect to the policy model parameters is calculated using the stochastic gradient ascent algorithm, and backpropagation is performed to update the policy model parameters.

7. A security operation system based on multi-agent collaboration, characterized in that, include: The edge sensing module, an edge sensing agent, intercepts the TLS record layer length sequence and message arrival time interval sequence of the encrypted session from the data service gateway; Based on the linear visibility geometry criterion, the time interval sequence is mapped to a temporal topological adjacency matrix. The attention mechanism is used to weight the length sequence and the matrix to generate a spatiotemporal topological embedding vector. The traffic shaping module maps the embedded vector input strategy model to multidimensional Gaussian distribution parameters, uses the traffic agent to sample and generate token adjustment coefficients, uses the adjustment coefficients to update the token delivery frequency of the gateway token bucket algorithm, calculates the target release timestamp of the data packets to be sent based on the current token status in the bucket, performs asynchronous delay operations on data packets that have not reached the timestamp, outputs the response traffic stream, and records the gateway processing delay index introduced by the asynchronous delay. The adversarial assessment module uses a random forest classifier to perform business fingerprinting on the response traffic flow and generate a traffic obfuscation escape rate. The optimization module reads the latency metric and escape rate to calculate the generalized advantage function estimate and the action probability ratio of the old and new strategies. It constructs an alternative objective function using a preset cutoff threshold and adjusts the parameters of the strategy model through a stochastic gradient ascent algorithm to maximize the alternative objective function.

Images