Encrypted traffic analysis method based on interaction spatio-temporal characteristics

By constructing the FITDect model, the problem of cryptographic traffic analysis dependence on shallow statistical features in the existing technology is solved, and the fusion analysis of the spatiotemporal characteristics of encrypted traffic is realized, and the ability to identify hidden behaviors of encrypted traffic is improved.

CN120378219AActive Publication Date: 2025-07-25INFORMATION & COMMNUNICATION BRANCH STATE GRID JIANGXI ELECTRIC POWER CO

Patent Information

Application Number
CN202510828318.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-06-20
Publication Date
2025-07-25
Estimated Expiration
2045-06-20

AI Technical Summary

Technical Problem

The existing deep flow detection technology relies on shallow statistical features for the analysis of encrypted traffic, making it difficult to effectively characterize the potential interaction characteristics between data packets within the flow, resulting in insufficient ability to identify hidden behaviors of encrypted traffic.

Method used

The FITDect model is constructed, the original encrypted stream is preprocessed through the input layer, the dynamic traffic interaction diagram is constructed, the GNN layer is used to analyze the spatiotemporal coupling relationship, and the classification decision is made in combination with the MLP layer to realize the fusion analysis of the spatiotemporal characteristics of the encrypted traffic.

Benefits of technology

It significantly enhances the ability to characterize the hidden behavior of encrypted traffic, can effectively capture the implicit behavior patterns of encrypted traffic, and provides a new method of malicious encrypted traffic detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120378219A_ABST
    Figure CN120378219A_ABST
Patent Text Reader

Abstract

The invention provides an encrypted traffic analysis method based on interaction spatio-temporal characteristics, relates to the field of network security, and aims at an original encrypted stream to construct a FITDect model consisting of an input layer, a GNN layer, an MLP layer and an output layer, and dynamically characterizes the traffic interaction diagram by mining the spatio-temporal interaction relationship of data packets in the original encrypted stream, constructing a dynamic traffic interaction diagram, and finally, analyzing the encrypted traffic. And analyzing a space-time coupling relationship through a GNN layer of the FITDect model by utilizing layered feature extraction, performing classification decision by utilizing an MLP layer of the FITDect model, sending a classification result to an output layer, and finally outputting an analysis result. According to the method, the problems that an existing deep flow detection technology depends on shallow statistical characteristics and a traditional deep packet detection technology fails are solved, fusion analysis of encrypted flow spatio-temporal characteristics is realized, the characterization capability of a model on encrypted flow hidden behaviors is remarkably enhanced, and a hidden behavior mode of the encrypted flow can be effectively captured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of network security, and particularly to an encrypted traffic analysis method based on interactive spatio-temporal features. Background Art

[0002] Many malware or malicious users take advantage of the concealment provided by encryption protocols to hide the transmission of malware or remote control instructions, so as to reduce the probability of the attack being detected. Different from traditional plaintext traffic, due to the encryption operation performed on the application layer of data packets by encryption protocols (such as SSL / TLS, IPSec, etc.), the application layer plaintext is invisible, which brings great challenges to both the QoS (Quality of Service) evaluation on the business side and the IDR (Intrusion Detection and Recognition) on the security side.

[0003] Existing traffic analysis methods can all be classified as DPI (Deep Packet Inspection) or DFI (Deep Flow Inspection). The DPI technology takes data packets as the smallest analysis unit and realizes the parsing of the transmission path and basic attributes of data packets by extracting five-tuple information such as protocol type, source IP address, destination IP address, source port, and destination port. As an important evolution of DPI in the field of traffic detection, DFI elevates the analysis dimension to the flow level. This method extracts statistical features such as flow duration, packet length distribution, transmission interval, and traffic period, and combines the time series pattern of traffic behavior to construct a macroscopic channel feature map of traffic. Based on the analysis paradigm of streaming spatio-temporal features, it inherits the parsing ability of DPI for network behavior, and effectively avoids the interference of encrypted payloads on content parsing by analyzing the overall behavior pattern of traffic, providing a new technical path for encrypted traffic identification.

[0004] The problems of the existing technologies mainly involve: Although the DFI technology realizes the macroscopic classification of encrypted traffic through statistical features such as flow duration, packet length distribution, and transmission interval, it still regards data packets as independent events with discrete time series in essence and cannot effectively represent the potential interaction features between data packets within a flow.

[0005] Therefore, it is urgent to develop a solution to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide an encrypted traffic analysis method based on interactive spatio-temporal features, aiming at the problems of strong concealment of encrypted traffic features, dependence of existing deep flow detection technologies on shallow statistical features, and invalidation of traditional deep packet detection technologies.

[0007] A method for encrypted traffic analysis based on interactive spatio-temporal features provided by the present invention adopts the following technical solution: A method for encrypted traffic analysis based on interactive spatio-temporal features specifically includes the following steps: Based on the original encrypted flow, construct a FITDect model composed of an input layer, a GNN layer, an MLP layer, and an output layer; Dynamically characterize the traffic interaction graph through the input layer of the FITDect model; Through hierarchical feature extraction, use the GNN layer of the FITDect model to analyze the spatio-temporal coupling relationship of the traffic interaction graph; Use the MLP layer of the FITDect model for classification decision-making, and send the classified result to the output layer of the FITDect model to output the analysis result.

[0008] A method for encrypted traffic analysis based on interactive spatio-temporal features, facing the original encrypted flow, constructs a FITDect model composed of an input layer, a GNN layer, an MLP layer, and an output layer. By mining the spatio-temporal interaction relationship of internal data packets in the original encrypted flow, construct a dynamic traffic interaction graph, perform dynamic characterization of the traffic interaction graph, use hierarchical feature extraction, analyze the spatio-temporal coupling relationship through the GNN layer of the FITDect model, and use the MLP layer of the FITDect model for classification decision-making. Send the classification result to the output layer, and finally output the analysis result. The present invention realizes the fusion analysis of the spatio-temporal features of encrypted traffic by constructing the graph structure information and node attribute information of the traffic interaction graph, significantly enhances the characterization ability of the model for the hidden behavior of encrypted traffic, can effectively capture the implicit behavior patterns of encrypted traffic, and provides a new method for malicious encrypted traffic detection.

[0009] Optionally, the input layer of the FITDect model preprocesses the original traffic data and constructs a graph structure as the input of the GNN layer of the FITDect model; the dynamic characterization of the traffic interaction graph through the input layer of the FITDect model is further expressed as:

[0010] Extract attack type traffic samples from the dataset for dataset sampling; re-represent the labels of different flows in the dataset; clear the invalid flows in the dataset; define interaction actions, and construct a traffic interaction graph according to the defined interaction actions; dynamically characterize the interaction features through the traffic interaction graph.

[0011] Optionally, the definition of the interaction action is further expressed as: Define a set of data packets with the same flow ID as Flow. The flow ID is a five-tuple consisting of a source IP address, a destination IP address, a source port number, a destination port number, and a protocol type. The sequence of interaction action directions is represented by the flow ID and the data packet set Flow.

[0012] Optionally, the construction of the traffic interaction graph is further represented as: Group the flows in chronological order to form a data packet sequence flow; abstract the sequence flow into several interaction actions and divide action slices; internally connect the nodes belonging to the same action slice; externally connect the nodes of adjacent action slices.

[0013] Optionally, the invalid flow is a flow with a data packet length less than 5 or a flow without a complete TCP handshake process.

[0014] Optionally, the hierarchical feature extraction is further represented as: Extract direction features and length features from the data packets of the flow; extract interaction features of independent communication behaviors; extract interaction state features.

[0015] Optionally, the GNN layer of the FITDect model is used to analyze the spatio-temporal coupling relationship of the traffic interaction graph, which is further represented as; Optionally, the GNN layer of the FITDect model aggregates and extracts the interaction features of the traffic interaction graph; the GNN layer of the FITDect model aggregates and extracts the node attribute features of the traffic interaction graph; analyze the spatio-temporal coupling relationship of the traffic interaction graph and generate an Embedding vector.

[0016] Optionally, the GraphSAGE algorithm is used to analyze the spatio-temporal coupling relationship of the traffic interaction graph. The GraphSAGE algorithm generates node embeddings by sampling and aggregating neighbor node information.

[0017] Optionally, the generation of an Embedding vector includes adopting a weighted average aggregation strategy, adjusting the influence of different nodes by introducing weights; calculating the weighted average of all node vectors after weighting the traffic interaction graph to generate a graph-level Embedding vector.

[0018] Optionally, the MLP layer of the FITDect model consists of 3 fully connected layers, which perform non-linear transformation on the node embeddings, and then capture feature combinations and patterns for classification decisions; the output layer of the FITDect model maps the classification results after being processed by the MLP layer to the task space and outputs the analysis results through a Softmax activator.

[0019] The beneficial effects of the present invention are as follows: A method for encrypted traffic analysis based on interactive spatio-temporal features, facing the original encrypted flow, constructs a FITDect model composed of an input layer, a GNN layer, an MLP layer, and an output layer. By mining the spatio-temporal interaction relationship of internal data packets in the original encrypted flow, constructs a dynamic traffic interaction graph, conducts dynamic characterization of the traffic interaction graph, uses hierarchical feature extraction, analyzes the spatio-temporal coupling relationship through the GNN layer of the FITDect model, and uses the MLP layer of the FITDect model for classification decision-making. Sends the classification result to the output layer, and finally outputs the analysis result. The present invention realizes the fusion analysis of the spatio-temporal features of encrypted traffic by constructing the graph structure information and node attribute information of the zero-traffic interaction graph, significantly enhances the characterization ability of the model for the hidden behavior of encrypted traffic, can effectively capture the implicit behavior patterns of encrypted traffic, and provides a new method for malicious encrypted traffic detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the FITDect model framework for the encrypted traffic analysis method of the present invention; Figure 2 It is a schematic diagram of the action slice division for the encrypted traffic analysis method of the present invention; Figure 3 It is a schematic diagram of the framework for constructing a traffic interaction graph for the encrypted traffic analysis method of the present invention; Figure 4 It is a schematic diagram of the encrypted traffic communication process for the encrypted traffic analysis method of the present invention; Figure 5 It is a schematic diagram of the interaction behavior for the encrypted traffic analysis method of the present invention; Figure 6 It is a schematic diagram of the encrypted analysis process for the encrypted traffic analysis method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein shall have the ordinary meaning as understood by those of ordinary skill in the art in the field to which the present invention belongs. The words such as "including" used herein are intended to mean that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0022] The embodiments of the present invention provide a method for encrypted traffic analysis based on interactive spatio-temporal features, as Figure 6As shown in the figure, it includes the following steps: S1. Based on the original encrypted stream, construct a FITDect model composed of an input layer, a GNN layer, an MLP layer, and an output layer; S2. Dynamically characterize the traffic interaction graph through the input layer of the FITDect model; S3. Through hierarchical feature extraction, use the GNN layer of the FITDect model to analyze the spatio-temporal coupling relationship of the traffic interaction graph; S4. Use the MLP layer of the FITDect model to make classification decisions, and send the classified results to the output layer of the FITDect model to output the analysis results.

[0023] A method for encrypted traffic analysis based on interactive spatio-temporal features, facing the original encrypted stream, constructs a FITDect model composed of an input layer, a GNN layer, an MLP layer, and an output layer. By mining the spatio-temporal interaction relationship between internal data packets in the original encrypted stream, constructs a dynamic traffic interaction graph, conducts dynamic characterization of the traffic interaction graph, uses hierarchical feature extraction, analyzes the spatio-temporal coupling relationship through the GNN layer of the FITDect model, and uses the MLP layer of the FITDect model to make classification decisions. Sends the classification result to the output layer, and finally outputs the analysis result. The present invention realizes the fusion analysis of the spatio-temporal features of encrypted traffic by constructing the graph structure information and node attribute information of the traffic interaction graph, significantly enhances the characterization ability of the model for the hidden behavior of encrypted traffic, can effectively capture the implicit behavior patterns of encrypted traffic, and provides a new method for malicious encrypted traffic detection.

[0024] In the context of the problem that the traditional DFI method has a shallow dependence on encrypted traffic analysis, the FITDect model proposed by the present invention uses a graph model to capture and represent the internal interaction patterns between data packets in encrypted traffic.

[0025] Specifically, in some embodiments, as Figure 1 shown, the FITDect model includes an input layer, a GNN layer, an MLP layer, and an output layer. The input layer first preprocesses the original traffic data, screens out invalid flows or overly short flows, and then divides the interaction actions and constructs a homogeneous graph as the input of the GNN layer.

[0026] The core objective of the preprocessing stage is to improve data quality. Raw traffic data often contains a large amount of noise, such as invalid flows caused by network jitter or device anomalies, and overly short flows due to transmission interruptions or acquisition errors. If this invalid data is directly input into the model, it will not only consume computing resources but may also interfere with the model's extraction of normal traffic characteristics. By setting thresholds such as traffic duration and packet integrity, the system can accurately filter out invalid flows, ensuring that subsequent analysis focuses on traffic segments with practical significance. This step is similar to the denoising operation in data cleaning, which can significantly reduce the interference of noise on model training and improve the overall usability of the dataset.

[0027] The process of partitioning interaction actions and constructing a homogeneous graph is essentially a structured modeling of traffic data. Traffic data is essentially a collection of nodes (such as IP addresses, device IDs) and edges (such as communication records, interaction behaviors), but the original data often exists in an unstructured form. By decomposing traffic interactions into specific actions (such as HTTP requests, DNS queries) and constructing a homogeneous graph based on the interaction relationships between nodes, the system can transform complex traffic patterns into a graph structure. The advantage of this modeling method is that it can simultaneously capture the attribute characteristics of nodes (such as device type, IP ownership) and the structural characteristics of edges (such as interaction frequency, temporal relationship), thus providing richer information input for the GNN layer. For example, in social network analysis, the follow relationships between users can be abstracted into a graph structure, and the device communication relationships in traffic analysis can be treated analogously.

[0028] The GNN layer is used to aggregate and extract the interaction features of the flow graph and the attribute features of the nodes themselves, generate an Embedding vector for each graph, and then send the vector classification result to the output layer through the MLP layer (multi-layer perceptron). And finally, the prediction result is output through the Softmax activator.

[0029] Specifically, in some embodiments, as Figure 1 shown, the input layer of the FITDect model preprocesses the raw traffic data and constructs a graph structure as the input to the GNN layer of the FITDect model; the dynamic characterization of the traffic interaction graph through the input layer of the FITDect model is further expressed as: Extract attack type traffic samples from the dataset for dataset sampling; re-represent the labels of different flows in the dataset; remove invalid flows from the dataset; define interaction actions, and construct a traffic interaction graph according to the defined interaction actions; dynamically characterize the interaction features through the traffic interaction graph.

[0030] Specifically, in some embodiments, the training dataset sampling is to extract a sufficient number of traffic samples of attack types from a large-scale public dataset, such as SQL injection, brute-force attack, man-in-the-middle attack, etc. The purpose of this step is to enrich the types of attack traffic as much as possible while ensuring that the scale of the training dataset is not too large, so that the model can recognize more attack interaction patterns.

[0031] Label encoding is to re-represent the labels of different flows in the dataset. On the one hand, it converts the labels into numerical types to facilitate subsequent model learning and classifier prediction. On the other hand, it also converts the multi-classification task in the dataset into a binary classification task.

[0032] The invalid flow is a flow with a packet length less than 5 or a flow without a complete TCP handshake process.

[0033] The data cleaning step mainly clears the incomplete or non-training-significant flows in the dataset, that is, invalid flows. In the present invention, a flow with a packet length less than 5 or a flow without a complete TCP handshake process is regarded as an invalid flow, and this step will clear these flows.

[0034] Further, in some embodiments, the definition of interaction actions is further expressed as: Define a set of packets with the same flow ID as Flow. The flow ID is a five-tuple composed of source IP address, destination IP address, source port number, destination port number, and protocol type; through the flow ID and the packet set Flow, the sequence of interaction actions is represented.

[0035] Specifically, in order to expose the interaction patterns of flows in the graph as much as possible, it is necessary to analyze the interaction process of communication behaviors between individuals. In some embodiments, the encrypted flow under study is defined as a session, which refers to a set of packets containing several packets with the same flow ID ( ), as shown in the following formula: ; The Flow is a flow formed by a sequence of several packets Packet(N).

[0036] is a five-tuple, composed of source IP address ( ), destination IP address ( ), source port number ( ), destination port number ( ), and protocol type ( ), as shown in the following formula: ; Based on the above formula, for the definition of the data packet set Flow and the flow ID ( ), the sequence representation of the interaction action direction can be derived. The specific formula is: ; Among them, represents the sequence of interaction action directions. The direction of the action is denoted as , and the value range is {-1, 1}. A value of -1 indicates that the interaction action is in the downlink direction (server - client), and a value of 1 indicates that the interaction action is in the uplink direction (client - server). represents the interaction action direction of the i-th data packet.

[0037] Based on the sequence representation formula of the interaction action direction, to describe the direction representation of the interaction actions in the flow, combined with this representation, the flow interaction process can be further abstracted into a sequence of receive and send behaviors. The specific formula is: ; Among them, represents send, represents receive, represents the sequence of behaviors, represents the i-th behavior element in the behavior sequence.

[0038] In the present invention, the five-tuple is the core design of the flow ID, which realizes the precise identification of network interactions. The source IP address and the destination IP address in the five-tuple clarify the two parties of the communication. The source port number and the destination port number further refine the specific services or applications of the communication, while the protocol type defines the rules and standards followed by the communication. The combination of these five elements makes each Flow an independent and unique interaction instance in the network, capable of accurately distinguishing different communication sessions. For example, in a complex network environment, multiple devices may communicate with the same server simultaneously. However, through the differentiation of the five-tuple, the independent interaction processes between each device and the server can be clearly identified, avoiding data confusion and misjudgment. This precise identification provides a reliable basis for subsequent traffic analysis, enabling analysts to conduct in-depth research on specific interaction instances.

[0039] Defining a set of data packets with the same flow ID as a Flow realizes the logical aggregation of network interactions. In actual network transmissions, a complete interaction process often consists of multiple data packets, which are transmitted in the network according to a certain order and rules. By aggregating data packets with the same flow ID into a single Flow, these scattered data packets can be integrated into an organic whole, thus more intuitively reflecting the entire process of the interaction. This logical aggregation not only simplifies the complexity of the data but also enables analysts to grasp key information such as the scale, duration, and data volume of the interaction from a macroscopic perspective. For example, when analyzing the transmission process of a video stream, by aggregating data packets with the same flow ID, one can clearly see the start time, end time, and the size of the data transmitted in the video stream, thereby evaluating the transmission quality and user experience of the video stream.

[0040] In some embodiments, such as Figure 2 shown, the process of associating each behavior with the interaction process is called action slicing. An action slice contains the longest consecutive segment of actions with the same direction during the interaction process. When an action slice starts, its direction must be opposite to that of the previous action slice and continues until the direction of the next action slice changes. Except for the initial action slice and the final action slice, the action slices in the interaction process have the characteristic of alternating between the same internal direction and opposite directions.

[0041] Using the flow ID and the data packet set Flow to represent the sequence of interaction action directions provides rich semantic information for traffic analysis. The sequence of interaction action directions reflects the data flow direction between the two communication parties. By analyzing these sequences, one can understand the data transmission patterns, service call relationships, and potential abnormal behaviors in the network. For example, in normal network communication, the client sends request data packets to the server, and the server returns response data packets after receiving the requests. This request-response interaction pattern is common. However, if it is found through analysis that there are a large number of request data packets sent from the server to the client in a certain Flow, but the client does not have corresponding responses, this may imply abnormal behaviors in the network, such as the server being attacked or the client having vulnerabilities. By deeply analyzing the sequence of interaction action directions, these potential security threats can be discovered in a timely manner, and corresponding measures can be taken for prevention.

[0042] In some embodiments, the construction of the traffic interaction graph is further represented as: Group the flows in chronological order to form a data packet sequence flow; abstract the sequence flow into several interaction actions and divide action slices; internally connect the nodes belonging to the same action slice; externally connect the nodes of adjacent action slices.

[0043] Furthermore, according to the above definition of action slices, a single-stream interaction graph can be generated by arranging and combining multiple layers of action slices. Edges can be constructed between adjacent slices to represent the potential associations between adjacent actions. This can not only preserve the original time-series information but also integrate richer information. It provides an intuitive description of the spatial structure characteristics of the interaction process. Next, the construction method of the interactive behavior graph will be described in detail.

[0044] Specifically, in some embodiments, as Figure 3 shown, based on the analysis of the transition states between interaction actions, the construction of the graph structure is divided into four steps, as follows: Step1. Flow grouping step. In this stage, according to the response data packets are extracted and grouped in chronological order to form a data packet sequence .

[0045] Step2. Divide action slices. In this stage, first, the data packet sequence is abstracted into a series of interaction actions, and the attribute sequence of these interaction actions is represented as . Taking as the initial input, the interaction process is expressed as independent slices to represent a single communication behavior.

[0046] Step3. Intra-slice connection. In this stage, first, the interaction actions are mapped to the vertices of the graph. According to the above analysis, there is no transition between the interaction actions within a single slice. Therefore, we use edges to connect these interaction actions in chronological order. In particular, the attribute of the edge representing the unchanged transition state is marked as 0.

[0047] Step4. Inter-slice connection. In this stage, based on the transition state, each interaction action within a slice will be connected to all the interaction actions in the adjacent slices. This step represents the relationship between the interaction actions in different slices, and the attribute of the edge representing the changed transition state is marked as 1.

[0048] Through these steps, the construction of the interaction process graph structure of the encrypted flow is completed. It should be noted that the vertices in the graph correspond to the mapping of the interaction actions, and each edge represents the transition state between the interaction actions. Since the transition state between the interaction actions represents a two-way process, the edges of the graph are undirected.

[0049] In the actual graph construction process, as Figure 4As shown, according to the interaction process of a given encrypted stream, an interaction timing diagram between a client and a server is presented. The diagram shows the interaction process between the two parties at different slices in chronological order. The entire interaction process is divided into six slices, namely Slice1 to Slice6, and each slice represents a specific stage or time period in the interaction process. This sliced interaction timing diagram can clearly show the interaction between the client and the server at different stages, including information such as the time points of interaction and the possible amount of data involved, which helps to analyze the timing characteristics, response time, and data transmission of the interaction between the two parties. It has important reference value for understanding the network interaction process, troubleshooting interaction problems, and optimizing interaction performance, etc.

[0050] As Figure 5 shown, an interaction behavior diagram of a multi-slice network structure is given. The diagram shows six slices, labeled Slice 1 to Slice 6 respectively. Each slice represents a local area in the network. The slices are connected to each other through nodes and edges, forming a complex network topology. In terms of the connection relationship between the slices, the nodes of different slices are associated with each other through edges, forming a cross-slice network structure. The node attributes in the graph structure consist of data packets and time, and at the same time, graph edge weights are introduced to further represent the static attention degree to the edges in different positions. There are connections between the nodes in Slice 2 and the nodes in Slice 3 and Slice 4. This cross-slice connection reflects the interaction and information transmission between different local network areas.

[0051] In the present invention, the core role of constructing the interactive behavior diagram is to reveal the internal structure of complex interaction relationships. In traffic analysis, data transmission in the network often involves dynamic interactions between multiple entities (such as IP addresses, devices, users, etc.). In the graph construction of the present invention, a data packet in the flow is regarded as a node, and the edge represents the association between data packets.

[0052] In the present invention, the use of the interaction behavior graph can effectively make up for the deficiencies of existing graph learning methods in encrypted traffic analysis methods. Existing graph learning methods have problems with insufficient granularity in interactive topology modeling. They usually only rely on external metadata such as IP address association or session similarity to construct the graph structure, but ignore the temporal interaction patterns at the packet level within encrypted traffic, such as the state transition of TLS handshake negotiation and the behavior sequence of encrypted payload fragmentation. The interaction behavior graph can penetrate into the encrypted traffic and capture these temporal interaction patterns at the packet level, thereby more accurately characterizing the cross-flow collaborative behavior in APT attacks. In the APT attack scenario, the activities of the attacker often involve multiple stages and the collaboration between different traffic flows. Ordinary graph learning methods are difficult to effectively identify such complex attack behaviors due to insufficient modeling granularity. The interaction behavior graph, with its ability to capture fine-grained interaction patterns, provides strong support for detecting such attacks.

[0053] At the same time, existing graph learning methods also have the problem of the disconnection between local and global features, and mostly adopt a single-level feature extraction strategy. Either focus on the statistical features of a single session or rely on the aggregation metrics of global traffic, and cannot achieve the joint modeling of fine-grained interaction patterns and macroscopic behavior patterns. Here, the fine-grained interaction patterns are, for example, the temporal dependence relationship of packets within a single flow, and the macroscopic behavior patterns are, for example, the topological relationship of instruction logic between multiple flows. This disconnection feature extraction method limits the detection sensitivity of the model to multi-stage encrypted attack chains. The interaction behavior graph can break this limitation. It can integrate fine-grained interaction patterns and macroscopic behavior patterns for modeling. By capturing fine-grained features such as the temporal dependence relationship of packets within a single flow and depicting macroscopic features such as the topological relationship of instruction logic between multiple flows, the interaction behavior graph can more comprehensively and accurately reflect the real behavior of network traffic, thereby enhancing the detection ability of multi-stage encrypted attack chains and timely discovering potential security threats.

[0054] Specifically, in some embodiments, the hierarchical feature extraction is further represented as: Extract the direction feature and length feature from the packets of the flow; extract the interaction feature of independent communication behavior; extract the interaction state feature.

[0055] Furthermore, after being processed by the input layer, a single flow is represented by a communication behavior graph, which is achieved based on the state transition of action slicing. Therefore, the communication behavior graph only contains the state transition information of action slicing, that is, the structural feature. However, in the action slicing conversion, the feature of the slice itself and the relationship feature between adjacent slices are also an important part of describing communication behavior. Therefore, the present invention proposes four interaction state features of action slicing, specifically: The internal state of the slice, including the direction and length extracted from the data packet, which are its basic features; the overall state of the slice, describing the characteristics of the action slice as an independent communication behavior in the interaction; the adjacent slice state, describing the context characteristics related to the action slice; the adjacent slice comparison, given in the form of a ratio, used to describe the change in the slice state.

[0056] Specifically, the interaction state features aim to present various feature information related to the interaction state, which are respectively composed of a feature name, a dimension, and a description. The feature name clearly gives different interaction state features, including length, direction, the number of slice actions, the ratio of the number of slice actions, slice length, the ratio of slice length, the ratio of action length, and the average number of bytes of slice actions. These features depict the interaction state from different perspectives.

[0057] The dimension clearly indicates the dimension value corresponding to each feature, and the value range varies from 1 to 3. The dimensions of length and direction are 1, the dimensions of the number of slice actions and slice length are 3, the dimensions of the ratio of the number of slice actions and the ratio of slice length are 2, and the dimensions of the ratio of action length and the average number of bytes of slice actions are 1. The setting of the dimension reflects the complexity of the feature in the data structure or the number of considerations. The description clearly explains the specific meaning of each feature. The length describes the length of each action; the direction represents the direction of each action; the number of slice actions refers to the specific number of actions included in the slice; the ratio of the number of slice actions reflects the proportional relationship of the number of interaction actions between the slice and its adjacent slice; the slice length represents the total length of the data packets of all actions in the slice; the ratio of slice length is the ratio of the total length of the action data packets between the slice and its adjacent slice; the ratio of action length represents the proportion of the length of a single action in the total length of the slice; and the average number of bytes of slice actions is the average length of the actions within a single slice.

[0058] Specifically, in some embodiments, the GNN layer of the FITDect model is used to analyze the spatio-temporal coupling relationship of the traffic interaction graph, which is further expressed as; The GNN layer of the FITDect model aggregates and extracts the interaction features of the traffic interaction graph; the GNN layer of the FITDect model aggregates and extracts the node attribute features of the traffic interaction graph; analyzes the spatio-temporal coupling relationship of the traffic interaction graph, and generates an Embedding vector.

[0059] In the present invention, the GNN layer aggregates and extracts the interaction features and node attribute features of the traffic interaction graph. This process can deeply mine the rich information contained in the traffic interaction graph. The interaction features reflect the patterns and rules of interactions between different nodes, such as the frequency and intensity of interactions, while the node attribute features characterize the characteristics of each node itself, such as the type and status of the node. Through the aggregation and extraction operations of the GNN layer, the model can more comprehensively understand the relationships and attributes between various elements in the traffic interaction graph, thereby providing a more accurate and richer basis for subsequent analysis and judgment.

[0060] Analyzing the spatio-temporal coupling relationship of the traffic interaction graph and generating an Embedding vector. This operation further improves the model's processing ability for traffic data. The spatio-temporal coupling relationship takes into account the associations of traffic interactions in both the time and space dimensions. In the time dimension, it can capture the changing trends of interactions over time, and in the space dimension, it can grasp the spatial layout and mutual influences between different nodes. Generating the Embedding vector represents the complex spatio-temporal coupling relationship obtained from the analysis in a concise and efficient vector form. This vector representation has many advantages. It is convenient for computers to process and calculate, can greatly reduce the dimension and complexity of data, and at the same time retains the key information of the traffic interaction graph. In subsequent tasks such as model training and prediction, the Embedding vector can be used as an input feature, enabling the model to better learn and identify the patterns and rules in traffic data, improving the accuracy and generalization ability of the model, and thus enhancing the performance and effect of the FITDect model in related application scenarios.

[0061] Furthermore, the GraphSAGE algorithm is used to analyze the spatio-temporal coupling relationship of the traffic interaction graph. The GraphSAGE algorithm generates node embeddings by sampling and aggregating neighbor node information.

[0062] Specifically, in order to effectively extract useful information from graph-structured data, the GraphSAGE algorithm based on sampling aggregation is used to generate graph embeddings. The GraphSAGE algorithm is suitable for medium and small graphs and also allows learning the embedding representation of unseen graph structures.

[0063] The GraphSAGE algorithm generates node embeddings by sampling and aggregating neighbor node information. This method not only reduces the computational complexity but also allows embedding learning for unseen nodes, but its performance highly depends on the choice of sampling strategy. In some embodiments, GraphSAGE will use uniform sampling, and the specific formula representation is: ; where represents the set of nodes sampled from the neighbor node set of node v, is a sampling function used to select a specified number of elements from a given set according to certain rules. is the number of samples per layer, which can be set in this article. v represents the central node currently being processed. This set represents all the nodes u in the graph that are connected to node v by an edge, that is, the direct neighbors of v; among them, is the set of all nodes in the graph.

[0064] In addition, the GCN algorithm can also be used to analyze the spatio-temporal coupling relationship of the traffic interaction graph.

[0065] The GCN algorithm learns node representations by applying convolutional operations on the adjacency matrix of the graph. This method can effectively capture local structural information, but its computational complexity is relatively high, especially when dealing with large-scale graphs, the efficiency is low. The formula for updating the node representation in the first layer of the GCN algorithm is: ; Among them, , is the adjacency matrix with self-connections added, is the adjacency matrix of the graph, is the identity matrix, is the degree matrix of, is the learnable weight matrix, is the feature matrix. At the same time, in some embodiments, the features of stacking two layers of convolution can be enhanced later, and the specific formula is expressed as: ; Among them, the Softmax function is used in the output layer for multi-classification problems, (adjacency matrix with self-connections added), and are both learnable weight matrices, is the feature matrix of the nodes in the graph, is the activation function, which is used to introduce non-linearity so that the model can learn more complex feature representations.

[0066] Furthermore, the GAT algorithm can also be used to analyze the spatio-temporal coupling relationship of the traffic interaction graph.

[0067] The GAT algorithm uses the attention mechanism to dynamically assign weights to the neighbors of each node, so as to better capture the important connection relationships in the graph. Although it is highly flexible and can handle complex graph structures, it may face the risk of overfitting during training.

[0068] In the present invention, the GraphSAGE algorithm can dynamically capture the spatio-temporal dependencies between network nodes, enhancing the ability to understand and predict complex interaction patterns. The core role of this step is to combine the spatial proximity and temporal evolution characteristics between nodes in the traffic interaction graph to construct an embedding representation that can reflect multi-dimensional associations, thereby providing a more expressive data basis for subsequent analysis.

[0069] In the traffic interaction graph, nodes usually represent network entities (such as servers, user terminals, or routers), and edges represent data transmission or interaction behaviors between entities. Traditional methods often analyze the spatial structure or time series in isolation and are difficult to capture the coupling effect between the two. During peak hours, the traffic surge of some nodes may be related not only to the load of their direct neighbors but also to the behaviors of remote nodes within the historical time window. Through the neighbor sampling and feature aggregation mechanism, GraphSAGE can dynamically integrate the information of multi-hop neighbors within a local range and, at the same time, combine the feature encoding in the time dimension to generate spatio-temporal aware embeddings of nodes. Such embeddings not only contain the static attributes of nodes but also incorporate the dynamic changes of their neighbors within the time window, enabling the model to understand the rule that "nodes that are spatially close exhibit similar behaviors within a specific time window".

[0070] In addition, the aggregation function of GraphSAGE is designed as a learnable module that can automatically adjust the feature combination method according to specific tasks. In the traffic interaction scenario, different types of data transmissions (such as video streams, file downloads) may have different sensitivities to spatio-temporal coupling relationships. By optimizing the aggregation function, the algorithm can more accurately capture these differences. For interactions with high real-time requirements, the algorithm may focus more on local features in the time dimension; while for long-term trend analysis, it may strengthen the integration of global information in the spatial dimension.

[0071] The advantages of the GraphSAGE algorithm lie in its flexibility and scalability. Through the sampling mechanism, the algorithm can handle large-scale traffic interaction graphs, avoiding the memory and computational overhead caused by full-graph calculation. At the same time, the diversity of aggregation functions (such as mean aggregation, LSTM aggregation, etc.) enables the algorithm to adapt to different types of spatio-temporal coupling patterns. For example, in the face of traffic interactions with periodic characteristics, LSTM aggregation can better capture long-term dependencies in the time series; while in scenarios where node features vary greatly, mean aggregation or pooling aggregation may be more efficient. In addition, the inductive learning ability of GraphSAGE enables the model to handle unseen nodes or subgraphs, which is particularly important for dynamically changing network environments. When new nodes are added or the behaviors of existing nodes change, the algorithm does not need to retrain the entire model, and only needs to generate new embedding representations based on local sampling and aggregation, thus greatly enhancing the practicality and adaptability of the model.

[0072] Specifically, in some embodiments, generating an Embedding vector includes adopting a weighted average aggregation strategy, adjusting the influence of different nodes by introducing weights; calculating the weighted average of all node vectors after weighting the traffic interaction graph to generate a graph-level Embedding vector.

[0073] In the present invention, the FITDect model adopts a weighted average aggregation strategy, applying different weights to the internal edges and external edges of the action slices to achieve the optimal effect. Compared with the four aggregation strategies of mean aggregation, maximum aggregation, weighted average aggregation, and LSTM aggregation, mean aggregation is simple and direct, suitable for capturing global features; maximum aggregation emphasizes highlighting the most important neighbor information, which helps to discover outliers or key nodes; weighted average aggregation combines the advantages of the former two, and can more flexibly adjust the influence of different neighbors by introducing weights, and is applicable to most scenarios; LSTM aggregation models the sequential pattern of node embeddings through the sequence model LSTM to generate the global embedding of the graph.

[0074] Generating a graph-level Embedding vector of the traffic interaction graph by adopting the weighted average aggregation strategy is to integrate node features through differential weight allocation, so as to capture the key information and overall pattern in the graph structure. The core role of this step is to aggregate the local features at the node level into a vector representation that can reflect the global characteristics, providing a basis for subsequent classification, prediction, or analysis tasks.

[0075] In the traffic interaction graph, there are usually significant differences in the importance of different nodes. Core routers or servers may carry a large amount of critical traffic, and their features have a much greater impact on the overall network state than edge nodes. Traditional unweighted average aggregation methods treat all nodes equally, resulting in the dilution of information of important nodes, while the noise of secondary nodes may mask key patterns. Weighted average aggregation can dynamically adjust the contribution of nodes to the final Embedding by introducing weights. Weights can be assigned according to indicators such as the degree centrality (number of connections), betweenness centrality (criticality of information transfer), or traffic load of nodes, so that the features of core nodes play a dominant role in the aggregation process. This mechanism enables the generated graph-level Embedding to more accurately reflect the core characteristics of the traffic interaction graph and avoid interference from irrelevant information.

[0076] Compared with the multi-round message passing in complex graph neural networks, weighted average only requires one aggregation operation to generate graph-level representations, significantly reducing the computational complexity. In real-time traffic monitoring scenarios, quickly generating graph-level Embeddings is crucial for timely response to network anomalies. Additionally, the weight assignment mechanism provides interpretability support. By analyzing the weight distribution, it is possible to intuitively identify which nodes have the greatest impact on the overall network state. Moreover, the interpretability of the weights makes the model more transparent. By analyzing the weight distribution, it is possible to intuitively identify which nodes contribute the most to the overall Embedding, providing a basis for network optimization or anomaly localization. Furthermore, the weight assignment mechanism of weighted average is inherently interpretable. By analyzing the correlation between weights and node attributes, it is possible to intuitively understand which factors (such as traffic load, number of connections) contribute the most to the graph-level Embedding.

[0077] The weighted average aggregation strategy is also robust to noise and outliers. In the traffic interaction graph, abnormal data of individual nodes (such as bursty traffic or measurement errors) may interfere with the global analysis. By reasonably designing the weights (such as confidence scores based on the historical behavior of nodes), the impact of abnormal nodes on the Embedding can be reduced. For nodes with frequently occurring abnormal traffic, their weights can be reduced or they can be directly excluded, thus ensuring the stability of the graph-level representation. This mechanism makes the algorithm more practical in real network environments.

[0078] Furthermore, in some embodiments, the MLP layer of the FITDect model consists of 3 fully connected layers, which perform a non-linear transformation on the node embeddings to capture feature combinations and patterns for classification decisions; the output layer of the FITDect model maps the classification results after being processed by the MLP layer to the task space and outputs the analysis results through a Softmax activator.

[0079] Specifically, the MLP layer of the FITDect model contains a three-layer neural network structure of input layer - hidden layer - output layer. The task type should be binary classification (flow malicious or benign), so in this layer, the softmax activation function will be used to compress the output to the interval (0, 1), representing the probability of the positive class. The three-layer forward propagation formula can be expressed as: ; where, , represents the input feature vector, that is, the graph Embedding vector, , , , represents the weight matrix of the fully connected layer; , , , represents the bias vector; is the ReLU activation function, defined as , and the ReLU activation function is used to introduce non-linearity, enabling the neural network to learn more complex feature representations; is the Sigmoid activation function, expressed as , which is used to describe the probability of the positive class.

[0080] For the loss function part, since the classification scenario is binary classification, binary cross-entropy is adopted here, and L2 regularization is additionally introduced to prevent overfitting. The loss function formula can be expressed as: ; where is the number of batch samples; , representing the true label, 0 represents normal traffic, and 1 represents abnormal traffic; , representing the predicted probability; represents the regularization coefficient, = 0.001; is the weight decay term.

[0081] In the present invention, the FITDect model architecture focuses on combining graph methods to extract interaction behavior features (spatiotemporal dimension) to solve encrypted traffic analysis. The FITDect model can analyze at a granularity level down to the flow. It focuses on the temporal features and payload interaction patterns within a single data flow; methods based on graph learning often need to aggregate multi-flow session information and are difficult to achieve real-time detection, while in-flow interaction analysis can complete traffic identification at the granularity of a single TCP / UDP flow. The FITDect model can also decouple the multi-dimensional spatiotemporal features of the flow for comprehensive analysis. The analysis method proposed in the present invention is an improvement on the basis of the traditional graph learning framework: by jointly modeling the graph structure information (spatial dimension) and node attribute information (time dimension) of the traffic interaction graph, the fusion analysis of the spatiotemporal features of encrypted traffic is realized. Compared with traditional graph learning methods that only rely on static topologies (such as fixed graphs based on IP relationships) or isolated node features (such as single-flow statistics), this spatiotemporal coupling modeling significantly enhances the model's ability to represent the hidden behaviors of encrypted traffic.

[0082] In some embodiments, based on three real encrypted traffic benchmark datasets, namely CIC-IDS 2017, CCCS-CIC-AndMal-2020, and MalwareTraffic Analysis, a dynamic interaction graph will be constructed. And through experiments, the best aggregation strategy of the GNN layer and the optimal GNN algorithm will be compared first. And finally, the traditional DFI method and recent GNN variant methods will be used as baseline algorithms to illustrate the model detection effect.

[0083] In some embodiments, running on a single desktop host, the hardware environment is a 16-core Intel(R) Xeon(R) Gold 6210R model CPU, an NVIDIA Geforce 3070TI model GPU, and 16GB of RAM memory. The software environment is the Ubuntu22.04 desktop operating system, the programming language Python 3.10.14, the deep learning framework Pytorch 2.2.2, and the graph neural network modeling library Torch-Geometric 2.5.3.

[0084] Specifically, in the embodiments, typical classification evaluation metrics in the fields of information retrieval and machine learning can be adopted, including accuracy, precision, recall, F1 score, and AUC score: ; ; ; ; ; Among them, TP is the true positive, representing the number of samples that are actually positive and are correctly predicted as positive; TN is the true negative, representing the number of samples that are actually negative and are correctly predicted as negative; FP is the false positive, representing the number of samples that are actually negative but are wrongly predicted as positive; FN is the false negative, representing the number of samples that are actually positive but are wrongly predicted as negative; the accuracy (AC) represents the ratio of the number of samples correctly classified by the model to the total number of samples, reflecting the overall correct rate of classification; the precision (PR) represents the ratio of the number of samples correctly classified as positive by the model to the total number of true positive samples, reflecting the completeness of the positive class coverage; represents the recall rate, indicating the proportion of samples that are actually positive and are correctly predicted as positive by the model, reflecting the ability of the model to find all positive class samples; the F1 value represents the harmonic mean of the precision and recall rates, comprehensively reflecting the precision and recall rates; the FPR represents the false positive rate, indicating the proportion of samples that are actually negative and are wrongly predicted as positive, which is often used for the drawing of the ROC curve; the AUC value refers to the area under the ROC curve, and the ROC curve shows the performance of the classifier by plotting the relationship between the true positive rate and the false positive rate at different thresholds.

[0085] To systematically analyze the impact of different graph learning algorithms on the model performance, three spatial domain graph neural network algorithms - GCN, GraphSAGE, and GAT - are respectively run, and their evaluation metrics are comprehensively measured.

[0086] According to the obtained F1-score comparison curve, it can be seen that the F1-score of GraphSAGE remains at a relatively high level from the beginning and converges basically after 20 rounds of training. For GCN and GAT, their F1-scores are unstable at lower rounds, and the convergence levels of both are not as high as that of GraphSAGE. Among them, the disadvantage of the GAT method is particularly obvious.

[0087] According to the obtained loss function curves on the training set and the test set, it can be seen that GraphSAGE also performs well in controlling the prediction error, with the fastest and most stable convergence speed. GCN also almost maintains a good prediction error, but there is a peak in the curve on the test set. The first 10 rounds of the GAT method have obvious fluctuations. Although it also converges roughly in the end, its stability still cannot be guaranteed.

[0088] Based on the above experiments, it can be determined that GraphSAGE is the best learning algorithm for the interaction behavior graph, which benefits from the good sampling and aggregation strategies of GraphSAGE. Specifically, the GraphSAGE sampling strategy adopted in the present invention is based on the idea of weighted average. For each node, taking the action slice where it is located as the anchor point, the information of all nodes is collected from the 3 slices above and below, and all are regarded as sampled nodes. At the same time, weights are assigned to the points to be sampled according to the slice distance. Under such a logic, the weights of the sampled points in adjacent slices are the highest, while the weights of the sampled points in separated slices decrease. In this way, the interaction features reflected by the context of the action slices are strengthened as much as possible to optimize the algorithm effect.

[0089] In the present invention, for the FITDect model, the best aggregation strategy for the interaction behavior graph is WAvg (weighted average aggregation), and the weight distribution should be more biased towards the nodes at the slice edges (in this paper, the distribution strategy of ɛ = 0.7 and ɵ = 0.3 is adopted). Compared with the GAT and GCN algorithms, GraphSAGE is more suitable for learning the interaction behavior graph constructed in this paper. GraphSAGE has the fastest learning speed and the most stable effect. This may be due to the multi-layer sampling and reasonable aggregation strategy of GraphSAGE, which can also predict well the graph patterns not seen in the training stage and has strong robustness. The FITDect model has achieved good results in the mixed dataset constructed in the present invention, and the comprehensive indicators of F1-score and AUC are equal to or better than those of traditional machine learning methods such as FlowLens and TSCRNN and new graph neural network methods such as GraphDApp and IBGC.

[0090] A method for encrypted traffic analysis based on interactive spatio-temporal features, which is oriented to the original encrypted flow, constructs a FITDect model consisting of an input layer, a GNN layer, an MLP layer and an output layer. By mining the spatio-temporal interaction relationship of the internal data packets of the original encrypted flow, a dynamic traffic interaction graph is constructed, and the dynamic characterization of the traffic interaction graph is carried out. Using hierarchical feature extraction, the spatio-temporal coupling relationship is analyzed through the GNN layer of the FITDect model, and the classification decision is made using the MLP layer of the FITDect model. The classification result is sent to the output layer, and finally the analysis result is output. The present invention realizes the fusion analysis of the spatio-temporal features of encrypted traffic by constructing the graph structure information and node attribute information of the zero-traffic interaction graph, significantly enhances the characterization ability of the model for the hidden behavior of encrypted traffic, can effectively capture the implicit behavior patterns of encrypted traffic, and provides a new method for malicious encrypted traffic detection.

[0091] Although the embodiments of the present invention have been described in detail above, it is obvious to those skilled in the art that various modifications and changes can be made to these embodiments. However, it should be understood that such modifications and changes are all within the scope and spirit of the present invention described in the claims. Moreover, the present invention described herein may have other embodiments and can be implemented or realized in various ways.

Claims

1. An encrypted traffic analysis method based on interactive spatio-temporal features, characterized in that, Including: Based on the original encrypted stream, construct a FITDect model composed of an input layer, a GNN layer, an MLP layer, and an output layer; Dynamically characterize the traffic interaction graph through the input layer of the FITDect model; Through hierarchical feature extraction, use the GNN layer of the FITDect model to analyze the spatio-temporal coupling relationship of the traffic interaction graph; Use the MLP layer of the FITDect model to make classification decisions, and send the classified results to the output layer of the FITDect model to output the analysis results.

2. The encryption traffic analysis method based on interactive spatio-temporal features according to claim 1, characterized in that The input layer of the FITDect model preprocesses the original traffic data and constructs a graph structure as the input of the GNN layer of the FITDect model; the dynamic characterization of the traffic interaction graph through the input layer of the FITDect model is further expressed as: Extract attack type traffic samples from the dataset and perform dataset sampling; Redefine the labels of different flows in the dataset; Remove invalid flows from the dataset; Define interaction actions, and construct a traffic interaction graph according to the defined interaction actions; Dynamically characterize the interaction features through the traffic interaction graph.

3. The encryption traffic analysis method based on interactive spatio-temporal features according to claim 2, wherein The definition of the interaction actions is further expressed as: Define a set of data packets with the same flow ID as Flow, and the flow ID is a five-tuple composed of a source IP address, a destination IP address, a source port number, a destination port number, and a protocol type; Represent the sequence of interaction action directions through the flow ID and the data packet set Flow.

4. The encryption traffic analysis method based on interactive spatio-temporal features according to claim 2, wherein The construction of the traffic interaction graph is further expressed as: Group the flows in chronological order to form a data packet sequence flow; Abstract the sequence flow into several interaction actions and divide action slices; Internally connect the nodes belonging to the same action slice; externally connect the nodes of adjacent action slices.

5. The encryption traffic analysis method based on interactive spatio-temporal features according to claim 2, wherein, The invalid flow is a flow with a data packet length less than 5 or a flow without a complete TCP handshake process.

6. The encryption traffic analysis method based on interactive spatio-temporal features according to claim 1, wherein, The hierarchical feature extraction is further expressed as: Extract direction features and length features from the data packets of the flow; Extract interaction features of independent communication behaviors; extract interaction status features.

7. A method for encrypted traffic analysis based on interactive spatio-temporal features according to claim 1, characterized in that, The use of the GNN layer of the FITDect model to analyze the spatio-temporal coupling relationship of the traffic interaction graph is further expressed as; The GNN layer of the FITDect model aggregates and extracts the interaction features of the traffic interaction graph; The GNN layer of the FITDect model aggregates and extracts the node attribute features of the traffic interaction graph; Analyze the spatio-temporal coupling relationship of the traffic interaction graph and generate an Embedding vector.

8. The encrypted traffic analysis method based on interactive spatio-temporal features according to claim 7, wherein, Use the GraphSAGE algorithm to analyze the spatio-temporal coupling relationship of the traffic interaction graph, and the GraphSAGE algorithm generates node embeddings by sampling and aggregating neighbor node information.

9. The encryption traffic analysis method based on interactive spatio-temporal features according to claim 7, wherein, The generation of an Embedding vector includes adopting a weighted average aggregation strategy and adjusting the influence of different nodes by introducing weights; Calculate the weighted average value of all node vectors after weighting the traffic interaction graph to generate a graph-level Embedding vector.

10. The encryption traffic analysis method based on interactive spatio-temporal features according to claim 1, wherein, The MLP layer of the FITDect model consists of three fully connected layers, which perform a non-linear transformation on the node embeddings to capture feature combinations and patterns for classification decisions; The output layer of the FITDect model maps the classification results processed by the MLP layer to the task space and outputs the analysis results through a Softmax activator.

Citation Information

Patent Citations

  • Opportunistic network link prediction method and device, and readable storage medium

    CN109347697A

  • Method for predicting time sequence network link by adopting GraphSAGE

    CN111756587A

  • D2D communication network slice allocation method based on deep reinforcement learning

    CN113163451A

  • Malicious encrypted traffic detection method based on spatial-temporal feature combination

    CN115174168A

  • Malicious encrypted traffic detection method based on graph convolutional network

    CN115174169A

Cited By

  • Internet of Things equipment binding method and system based on AI behavior data

    CN120785758A

  • Encrypted traffic data leakage traceability method and system based on spatio-temporal characteristics

    CN120915604A