A network security intelligent detection method and system based on traffic analysis
By collecting and analyzing intranet traffic data, generating mimicry decoys and constructing a phantom matrix, the problem of missed detection of unknown protocols and highly obfuscated payloads in existing technologies is solved, and accurate detection of risks to intranet hosts and isolation of attack behaviors are achieved.
Patent Information
- Application Number
- CN202611374514.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-07
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-07
AI Technical Summary
Existing technologies are insufficient to accurately characterize host risks, are easily detected by attackers, fail to detect behavioral mutations in host behavior graphs, and are prone to missing unknown protocols, encrypted or highly obfuscated payloads.
Collect bidirectional network traffic data from the internal network, analyze potential threat data through a dual-mode matching engine, establish an internal network behavior profile, generate mimicry decoys, dynamically deploy them to high-risk nodes, construct a phantom matrix to isolate attack behavior, and generate a network security detection report.
It achieves accurate detection of unknown protocols and highly obfuscated payloads, improves the ability to identify unknown threats and covert lateral movement, identifies online hosts and operating system assets on the internal network without active scanning, dynamically adjusts decoys to maintain deception effectiveness, and isolates attack behaviors.
Smart Images

Figure CN122870565B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cyberspace security technology, and in particular to an intelligent network security detection method and system based on traffic analysis. Background Technology
[0002] As advanced persistent threats continue to evolve, the cybersecurity risks faced by internal networks are becoming increasingly severe. These attacks are characterized by long-term infiltration and multi-stage penetration, often bypassing traditional passive defense systems based on boundary isolation and signature matching.
[0003] Currently, Chinese invention patent application number CN122394876A discloses a method, apparatus, and device for processing audit tasks of cybersecurity compliance documents. The method includes: acquiring file data and lineage metadata; storing the hash value of the file data on the blockchain to obtain a storage serial number; determining the nodes of the file lineage graph based on the storage serial number; determining the directed edges between nodes based on the lineage metadata; obtaining the file lineage graph based on the nodes and directed edges; monitoring changes to obtain file hash change events; analyzing the impact scope of the file lineage graph based on the storage serial number to obtain a list of storage serial numbers for downstream cybersecurity compliance documents; generating a re-audit task based on the storage serial number list, distributing it, and storing it on the blockchain; verifying the downstream cybersecurity compliance documents through a compliance rule smart contract engine, generating audit results, and storing the audit results on the blockchain.
[0004] The aforementioned technologies are insufficient to solve the detection problems of unknown protocols, encrypted or highly obfuscated payloads, and their bypass traffic acquisition and deep analysis capabilities are inadequate, making it easy to miss unknown protocols, encrypted or highly obfuscated payloads. Summary of the Invention
[0005] The technical problem solved by this invention is that existing technologies are difficult to accurately characterize host risks, are easily detected by attackers, and fail to detect behavioral gene mutations in host behavior maps.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a network security intelligent detection method based on traffic parsing includes the following steps: Step S1: Collect and parse the bidirectional network traffic data of the intranet to obtain traffic metadata. Analyze potential threat data through a dual-mode matching engine and extract traffic behavior features based on a segmentation algorithm. Step S2: Based on the traffic metadata and traffic behavior characteristics, automatically identify online hosts on the intranet, establish intranet behavior profiles for each host, and dynamically generate mimicry decoys; Step S3: Generate a dynamic attack graph based on potential threat data and intranet behavior profiles, and deploy the mimicry decoy to high-risk nodes in the intranet using a game theory model; Step S4: Continuously detect access traffic pointing to the mimic decoy, obtain network attack behavior, and use dynamic trapping technology to construct a phantom matrix and isolate the network attack behavior; Step S5: Save the movement trajectory of the network attack behavior and generate a network security detection report.
[0007] Preferably, step S1 includes the following sub-steps: Step S11: Capture host-to-host traffic data as first data, capture host-to-gateway traffic data as second data, and use the first data and the second data as bidirectional intranet traffic data. Step S12: Decompose and decode the intranet bidirectional traffic data, extract header information and session flow status of each layer and perform standardization processing to obtain traffic metadata, which includes a five-tuple, session traffic status, application layer attributes and protocol parsing status identifier. Step S13: Analyze the traffic metadata using a dual-mode matching engine to generate potential threat data, which includes threat matching results, rule detection hit characteristics, and malicious network characteristics. Step S14: From the intranet bidirectional network traffic data, remove traffic data with known risks by matching the protocol parsing status of the traffic metadata with the potential threat data, filter out the remaining traffic data, and call the segmentation algorithm to cut the remaining traffic data to obtain the corresponding functional fields; Step S15: By summarizing and extracting the functional fields and traffic metadata, traffic behavior characteristics are obtained. The traffic behavior characteristics include abnormal commands, API unauthorized calls, malicious payload distribution, communication frequency fluctuations, and session interaction relationships.
[0008] Preferably, the processing logic of the dual-mode matching engine is as follows: Extract the five-tuple and corresponding timestamp of the traffic metadata, and construct a host behavior graph based on a sliding time window with each host in the intranet as a node and communication behavior as an edge. The host behavior graph includes port distribution, set of communication peer IPs and active time periods. Obtain the metadata of the currently arriving traffic and extract the current communication behavior. Traverse the host behavior graph. When no sample in the host behavior graph has the same communication pattern as the current communication behavior, it is determined to be a behavioral gene mutation, and a rule is generated to detect the hit feature. Extract communication behaviors corresponding to behavioral gene mutations from multiple hosts, perform correlation analysis, and generate malicious network characteristics; Based on the rules, the system detects hit features and malicious network features, generates threat matching results, and packages them to generate potential threat data.
[0009] Preferably, step S2 includes the following sub-steps: Step S21: Extract the five-tuple and application layer attributes of the traffic metadata, and perform deduplication and aggregation using the source IP address and destination IP address as indexes to obtain the set of host IPs of all active communication in the current intranet. For each host in the set of host IPs, extract the corresponding MAC address, operating system type and open service list to generate a set of online hosts in the intranet. Step S22: Based on the traffic metadata and traffic behavior characteristics, a dual-granularity fusion modeling method is used to establish an intranet behavior profile for the set of intranet online hosts; Step S23: The behavior profile of the intranet is processed using a behavior dynamic mimicry engine to generate a mimicry decoy.
[0010] Preferably, steps S22 and S23 further include the following sub-steps: Step S221: Extract the session traffic status and communication timestamp of each host, and construct the vertical behavior sequence of each host on the time axis. The vertical behavior sequence is the port access frequency change, uplink and downlink traffic ratio fluctuation and session duration distribution of the current host in different time windows. Through the communication frequency anomaly and session interaction relationship in the traffic behavior characteristics, mark the abnormal time sequence nodes in the vertical behavior sequence of the current host. Step S222: Using all hosts in the set of online hosts in the intranet as nodes and the actual communication records between hosts as edges, extract the five-tuples from the traffic metadata to construct a full network communication topology graph. Calculate the topology centrality index of each host in the full network communication topology graph, and calculate the node value of each host in the intranet topology using the topology centrality index. The topology centrality index includes the number of communication connections, the frequency of communication relay nodes, and the average communication distance with other nodes. Step S223: The vertical behavior sequence of each host is fused with the node value to generate the intranet behavior profile of the current host. The intranet behavior profile includes host identity, open service list, vertical behavior sequence, node value and network topology location. Step S231: Traverse the intranet behavior profile, sort the node values in descending order, select the host with the highest node value as the target host, extract the open service list, port configuration and operating system type from the intranet behavior profile of the target host as a mimicry feature set, construct a simulation instance in the virtualization container based on the mimicry feature set, and complete the static feature cloning. Step S232: Continuously monitor the real-time traffic metadata of the target host and synchronize it to the corresponding simulation instance in real time; Step S233: Based on the abnormal instructions and malicious payload distribution in the traffic behavior characteristics, a capture probe is pre-set in the simulation instance, and the simulation instance and the corresponding mimicry feature set and capture probe configuration are encapsulated to generate a mimicry decoy.
[0011] Preferably, step S3 includes the following sub-steps: Step S31: Construct a dynamic attack map using the potential threat data and intranet behavior profile through the network-wide communication topology map; Step S32: Based on the dynamic attack graph, perform a probability assessment on the attack risk of each node, generate the node risk probability, and select a set of high-risk deployment nodes based on the node risk probability. Step S33: Based on the set of high-risk deployment nodes and node risk probabilities, extract attacker behavior types from traffic behavior characteristics, construct attacker type belief vectors, introduce business fluctuation disturbance factors, construct belief-driven dynamic game model, and calculate the optimal decoy deployment strategy. Step S34: According to the optimal decoy deployment strategy, deploy the mimicry decoy to the corresponding intranet location in the high-risk deployment node set to generate the decoy deployment topology.
[0012] Preferably, step S4 includes the following sub-steps: Step S41: Based on the decoy deployment topology, continuously monitor the access traffic of the mimic decoy, extract attack indicators and record them as network attack behaviors; Step S42: Infer intent from the attack indicators in the network attack behavior to generate attack tactical intent; Step S43: Based on the deception defense information entropy model, continuously calculate the entropy value of the current trapping scenario, evaluate the deception effectiveness of the mimicry decoy, and trigger the reconstruction of the phantom matrix when the entropy value is lower than the preset safety threshold. Step S44: Based on the attack tactical intent, dynamically construct a phantom matrix in conjunction with the intranet behavior profile; Step S45: Input all communication traffic corresponding to the network attack behavior into the phantom matrix to complete the isolation of the network attack behavior.
[0013] Preferably, step S42 specifically includes: Extract the operation instruction sequence and payload content from the network attack behavior, preset the proportion threshold, the proportion threshold includes a first proportion threshold, a second proportion threshold and a third proportion threshold, count the proportion of the occurrence frequency of each type of instruction in the operation instruction sequence to the sum of the occurrence frequencies of all instructions, and identify the attack stage, the attack stage includes a first stage, a second stage and a third stage. When the sum of the frequencies of port scanning commands and service probe commands is greater than a first ratio threshold, and the payload content does not contain vulnerability exploitation feature fields, the current attack stage is determined to be the first stage. When the sum of the frequencies of vulnerability exploitation commands, lateral scanning commands, and credential retrieval commands exceeds a second threshold, and the payload content contains vulnerability exploitation feature fields that match the open service list of the mimic decoy, the current attack phase is determined to be the second phase. When the frequency of data outbound transmission commands exceeds the third proportion threshold, and the length of a single payload in the payload content exceeds the preset payload length threshold, the current attack stage is determined to be the third stage. The attack phase determination results are structurally aggregated with the host nodes in the internal network behavior profile that have communication records with the attack source IP to generate attack tactical intent.
[0014] Preferably, step S5 includes the following sub-steps: Step S51: Integrate the network attack behavior, the decoy deployment topology, and the traffic metadata to reconstruct the attacker's complete operation sequence and generate an attack movement trajectory; Step S52: Retrieve the attack movement trajectory and the potential threat data, and output the security event record; Step S53: Feed back the attack tactical intent to step S13 and update the host behavior graph; Step S54: Summarize the potential threat data, attack movement trajectory, and attack tactical intent to generate a network security detection report.
[0015] Secondly, a network security intelligent detection system based on traffic analysis includes: The traffic awareness module is used to parse the obtained traffic metadata, analyze potential threat data, and extract traffic behavior features based on a segmentation algorithm. The profile generation module is used to identify online hosts on the intranet through traffic metadata and traffic behavior characteristics, and to build intranet behavior profiles and dynamically generate mimicry decoys; The attack game module is used to generate dynamic attack graphs based on potential threat data and internal network behavior profiles, and to deploy mimicry decoys to high-risk nodes in the internal network. The decoy isolation module is used to detect access traffic directed to the mimicry decoy, obtain network attack behavior, construct a phantom matrix, and isolate network attack behavior; The attack tracing module is used to save the movement trajectory of the network attack behavior and generate a network security detection report. The beneficial effects of this invention are as follows: This invention collects bidirectional traffic from host to host (east-west) and from host to gateway (north-south) at the intranet boundary bypass location, covering intranet lateral movement and intranet-extranet communication channels. It does not actively send packets, does not occupy service bandwidth, and does not change the original network topology, having no impact on the operation of core businesses such as OA, business systems, and databases. It introduces protocol parsing status identifiers, retaining unparsed payloads and sending them to a segmentation algorithm for processing. Through a self-supervised model and lateral inference and vertical correction, it performs field segmentation and function identification on unknown protocol payloads, extracting traffic behavior characteristics such as abnormal instructions, API unauthorized calls, malicious payload distribution, communication frequency anomalies, and session interaction relationships, effectively improving the detection capability of highly obfuscated, privatized, or encrypted APT communications. It performs behavioral gene mutation detection based on host behavior graphs, generating rules to detect hit features through multi-dimensional deviation judgments such as port distribution, communication peer IP sets, and active time periods. It constructs a mutation propagation topology from a network-wide perspective, correlates and analyzes cross-host collaborative attack behaviors, and generates malicious network characteristics. This mechanism does not rely on a single signature and can detect unknown threats and covert lateral movements that deviate from normal behavior patterns. It passively identifies intranet online hosts, operating systems, open services, MAC addresses, and other asset information based on traffic metadata, without requiring active scanning. Through vertical behavioral sequence modeling and network-wide communication topology node value calculation, it achieves a dual-granularity fusion profile of host behavior in both time and topology dimensions, providing accurate basis for vulnerable node screening, high-value target identification, and decoy generation. It continuously monitors access traffic directed to the mimicry decoy, extracts attack indicators, and infers attack tactical intentions. Based on a deception defense information entropy model, it evaluates the deception effectiveness of the trapping scenario. When the entropy value decreases, it automatically injects new decoys or reconstructs the characteristics of existing decoys, forming a phantom matrix. Through transparent proxies, attack traffic is diverted to the phantom matrix, isolating attack behavior, protecting real assets, and simultaneously invalidating the attacker's accumulated knowledge, restoring deception effectiveness. Attached Figure Description
[0016] Figure 1 A flowchart illustrating the steps of an intelligent network security detection method based on traffic parsing, as provided in one embodiment of the present invention; Figure 2 This is a basic flowchart of a network security intelligent detection system based on traffic parsing, provided as an embodiment of the present invention. Detailed Implementation
[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0018] Example 1, referring to Figure 1This paper provides an intelligent network security detection method based on traffic parsing, which includes the following steps: Step S1: Collect and parse the bidirectional network traffic data of the intranet to obtain traffic metadata. Analyze potential threat data through a dual-mode matching engine and extract traffic behavior features based on a segmentation algorithm. Step S2: Automatically identify online hosts on the intranet based on traffic metadata and traffic behavior characteristics, establish intranet behavior profiles for each host, and dynamically generate mimicry decoys; Step S3: Generate a dynamic attack graph based on potential threat data and intranet behavior profiles, and deploy the mimicry decoy to high-risk nodes in the intranet through a game theory model; Step S4: Continuously detect access traffic pointing to the mimic decoy to obtain network attack behavior, and use dynamic trapping technology to construct a phantom matrix and isolate network attack behavior; Step S5: Save the movement trajectory of the network attack behavior and generate a network security detection report.
[0019] In this embodiment, by collecting first host-to-host data and second host-to-gateway data at the boundary bypass location of the government and enterprise intranet, full coverage of east-west lateral traffic and north-south outbound traffic is achieved. The entire process involves no active packet generation, no consumption of service bandwidth, and zero interference with the original network topology and service operation. During traffic parsing, a protocol parsing status identifier is introduced to mark and retain unknown payloads that cannot be identified by standard protocol parsers. Combined with a segmentation algorithm, unparsed payloads undergo self-supervised field segmentation and function identification. This enables the extraction of abnormal commands, malicious payload distribution, communication frequency anomalies, and session interaction relationships from highly obfuscated, privatized, or encrypted APT communications. This approach effectively overcomes the problem of traditional traffic detection tools missing unknown protocols. The dual-mode matching engine uses host behavior graphs as a basis for detecting behavioral mutations. It generates rules to detect hit features by analyzing deviations from port distribution, communication peer IP sets, and active time periods. Furthermore, it constructs a mutation propagation topology from a network-wide perspective for cross-host correlation analysis, generating malicious network characteristics. This mechanism does not rely on a single feature signature and can identify known and unknown threats that deviate from normal behavior patterns, exhibiting higher detection sensitivity for behaviors such as lateral movement within the internal network, low-frequency infiltration, and covert external connections. It constructs an internal network behavior profile through dual-granularity fusion modeling, establishing a vertical behavior sequence on the timeline and... The system calculates node value based on the entire network communication topology, making asset identification, behavioral baseline characterization, and high-value target screening more accurate. Based on this, it dynamically generates mimicry decoys, using static feature cloning and real-time traffic synchronization to maintain high fidelity and dynamic consistency between the decoys and real assets, reducing the risk of the decoys being identified by attackers. The system optimizes mimicry decoy deployment based on dynamic attack graphs and belief-driven dynamic game models: it assesses node risk through Bayesian probability inference and filters high-risk deployment nodes; it introduces attacker type belief vectors and business fluctuation disturbance factors to make the defense strategy more closely resemble real attack and defense scenarios, optimizing deployment location and type under limited decoy resources; and it continuously monitors targets that mimicry. The system extracts attack indicators from the access traffic of decoys and infers the three-stage attack tactical intent. Based on the deception defense information entropy model, it dynamically evaluates the deception effectiveness. When the entropy value is lower than the security threshold, it triggers the reconstruction of the phantom matrix and isolates the attack behavior within the phantom matrix through a transparent proxy, ensuring that real internal network assets are not affected. At the same time, it continuously consumes the attacker's resources and restores the deception effectiveness of the trapping scenario. It integrates network attack behavior, decoy deployment topology, and traffic metadata to restore the attacker's complete movement trajectory and generate a standardized network security detection report. At the same time, it feeds back new attack techniques captured during the trapping process to update the host behavior graph, significantly improving the sustainability and adaptability of internal network security detection and proactive defense.
[0020] Step S1 includes the following sub-steps: Step S11: Capture host-to-host traffic data as first data, capture host-to-gateway traffic data as second data, and use the first data and the second data as bidirectional intranet traffic data. Step S12: Decompose and decode the bidirectional traffic data of the intranet, extract the header information and session flow status of each layer and perform standardization processing to obtain traffic metadata, which includes a five-tuple, session traffic status, application layer attributes and protocol parsing status identifier. Step S13: Analyze traffic metadata using a dual-mode matching engine to generate potential threat data, which includes threat matching results, rule detection hit characteristics, and malicious network characteristics. Step S14: From the intranet bidirectional traffic data, perform a dual screening based on the protocol parsing status identifier of the traffic metadata and the threat matching results of the potential threat data: Traffic data with known risk status in threat matching results are removed, as are traffic data with protocol parsing status marked as parsed. Traffic data that simultaneously meet the conditions of unparsed protocol parsing status and no risk status in threat matching results are filtered as the remaining traffic data. The segmentation algorithm is then used to cut the remaining traffic data to obtain the corresponding functional fields. Step S15: By summarizing and extracting functional fields and traffic metadata, traffic behavior characteristics are obtained. These characteristics include abnormal commands, API unauthorized calls, malicious payload distribution, communication frequency fluctuations, and session interaction relationships.
[0021] In this embodiment, a traffic acquisition module deployed in a bypass manner captures host-to-host traffic data in the intranet in real time as the first data, and host-to-gateway traffic data as the second data, without actively sending packets or occupying service bandwidth. The first and second data are then combined into bidirectional intranet traffic data. This is mainly achieved by accessing the intranet through the mirror port of the core switch or a network splitter, diverting a complete copy of the intranet's entire traffic to the acquisition module, ensuring that the original data link remains unaffected. This process ensures complete coverage of the intranet's entire traffic. The first data covers east-west traffic between hosts, which is the main carrier during the lateral movement phase of APT attacks. The second data covers north-south traffic between the intranet and external networks, which is the main channel for Trojan horse backlinks, data leaks, and other such activities.
[0022] Taking a real-world deployment scenario as an example, when an SMB protocol communication occurs between internal network host A (IP: 192.168.1.10) and host B (IP: 192.168.1.20), a complete copy of the data frame of this communication is collected as the first data. When host A initiates an HTTP request to the external network gateway (IP: 192.168.1.1), a complete copy of the data frame of this request is collected as the second data. The two types of data are merged to form complete bidirectional traffic data of the internal network, ensuring that both east-west horizontal movement traffic and north-south external connection traffic are within the collection range.
[0023] When performing multi-layer protocol stripping on bidirectional traffic data within the intranet, taking a typical abnormal session as an example: When intranet host A (192.168.1.10) initiates a connection to the non-standard port 8443 of the external IP (203.0.113.10), S12 sequentially strips the Ethernet frame header, IP header, and TCP header, extracting the five-tuple as (192.168.1.10, 203.0.113.10, 54321, 8443, TCP). The session traffic status record duration is 1200 seconds, the number of uplink packets is 320, and the total payload length is 4.8MB. The application layer attempts to decode the payload using standard protocol parsers such as HTTP, TLS, and SSH, but all attempts fail. Therefore, the protocol parsing status flag is set to unparsed. The session traffic metadata is retained and transferred to S13 for dual-mode matching detection.
[0024] The protocol parsing status identifier is a key design feature of S12. Traditional traffic parsing tools discard unidentifiable payloads directly, while this solution uses this identifier to mark and retain unknown payloads, providing a clear processing entry point for the S14 segmentation algorithm and avoiding missed detection of highly obfuscated APT private protocol traffic. In this embodiment, the dual-mode matching engine performs behavioral gene mutation detection on the traffic metadata output by S12, demonstrating the generation process of potential threat data in two typical scenarios: Scenario 1 (Rule Generation Detection Hit Features): The host behavior graph of host B (192.168.1.20) shows that its historical active periods are all from 9:00 to 18:00. Port 445 has never appeared in the usual destination port distribution, and 192.168.1.80 is not included in the communication peer IP set. In a certain time window of 2:00 AM, it was detected that host B initiated an SMB connection to port 445 of 192.168.1.80, with a total load of 3.2MB. The time period deviation event and port deviation event were triggered simultaneously, and it was determined that a behavioral gene mutation had occurred. The source IP (192.168.1.20), destination IP (192.168.1.80), port (445), and timestamp were extracted to generate rule detection hit features. The threat matching result was configured as a known risk status, and this traffic data was marked and removed in S14. Scenario 2 (no features generated): Host A (192.168.1.10) initiates a connection to port 8443 of external IP (203.0.113.10). The comparison of its traffic metadata with the host behavior graph does not trigger any deviation event (this host has accessed port 8443 in the past). The collision with the external threat intelligence database is not a match. The threat matching result is configured as a risk-free state, but its protocol resolution status is marked as unresolved. This traffic will be processed as residual traffic data in S14 and entered into the segmentation algorithm.
[0025] Taking the session in which host A initiates a connection to port 8443 of 203.0.113.10 in scenario two as an example, the process of filtering and segmenting the remaining traffic data is explained: This session meets both screening criteria: the protocol parsing status is unparsed and the threat matching result is risk-free. Therefore, it is identified as remaining traffic data and sent to the segmentation algorithm for processing.
[0026] The segmentation algorithm processes the session payload as follows: The self-supervised model calculates the correlation degree between adjacent bytes in the payload byte sequence, generates a byte relationship curve, and discovers the following correlation degree distribution pattern: Byte positions 0-1 (0xAB 0xCD): extremely high correlation (0.97), belonging to the same field; Byte positions 1-2 (0xCD to 0x00): The correlation drops sharply to 0.12, indicating a field boundary; Byte positions 2-3 (0x00 0x12): High correlation (0.89), within the same field; Byte positions 3-4 (0x12 to 0x03): The correlation drops sharply to 0.08, indicating a field boundary; Horizontal reasoning involves segmenting at points where the relevance drops sharply, initially dividing candidate fields into [0xAB 0xCD], [0x00 0x12], [0x03], and [subsequent payload data]. Vertical correction involves comparing multiple messages from the same session to verify that [0xAB 0xCD] remains constant across all messages (protocol identifier field), [0x03] changes frequently with session content (instruction control code field), and the length of subsequent payload data is variable (payload data field). Finally, the segmentation boundary is confirmed, and corresponding functional fields are generated.
[0027] Continuing with the functional fields of the host A session mentioned above as input, the process of extracting traffic behavior characteristics is explained: The value 0x03 in the instruction control code field corresponds to a download execution instruction, which has no corresponding semantics in any known legitimate business protocol, and is extracted as an abnormal instruction feature. The PE file header feature (MZ identifier) was identified in the payload data field, confirming that the payload is an executable file, and is extracted as a malicious payload distribution feature. The session traffic status in the traffic metadata shows that the session initiates communication once every 300 seconds, which is significantly inconsistent with the Poisson distribution pattern of the historical communication frequency of host A, and is extracted as a communication frequency anomaly feature (typical heartbeat back connection feature). The time-series interaction sequence in the function field shows that the latency variance of the request and response is extremely small (standard deviation is only 12ms), presenting a typical automated C2 response feature, and is extracted as a session interaction relationship feature. No unauthorized interface call records were found, and the API unauthorized call feature for this case is empty.
[0028] The processing logic of the dual-mode matching engine is as follows: Extract the five-tuple of traffic metadata and the corresponding timestamp. Using each host in the intranet as a node and communication behavior as an edge, construct a host behavior graph based on a sliding time window. The host behavior graph includes port distribution, set of communication peer IPs, and active time periods. Obtain the metadata of the currently arriving traffic and extract the current communication behavior. Traverse the host behavior graph. If no sample in the host behavior graph has the same communication pattern as the current communication behavior, it is determined to be a behavioral gene mutation. Generate rules to detect the hit features. Extract communication behaviors corresponding to behavioral gene mutations from multiple hosts, perform correlation analysis, and generate malicious network characteristics; Based on the rules, the system detects hit characteristics and malicious network characteristics, generates threat matching results, and packages them to generate potential threat data.
[0029] In this embodiment, the traffic metadata is sliced with a preset time window length and step size. Within each time window, communication behavior samples of each host node are collected. Within each time window, for each host node, communication pattern features are extracted from the traffic metadata and used as edges of the graph. The communication pattern features include the distribution of commonly used destination ports, the set of communication peer IPs, and the distribution of active time periods. The form of the distribution of commonly used destination ports is: when host i is the communication initiator, the frequency distribution vector of its destination ports. ,in For port number, For access frequency, the communication peer IP set is the historical communication peer IP address set of the current host and the communication frequency weight of each peer. The active time period distribution in this embodiment is to count the frequency of communication initiated by the host in each time period of a day in hours, and generate a 24-dimensional active time vector. The communication pattern features extracted in multiple consecutive time windows are averaged. When the feature change amplitude between two adjacent windows is lower than the preset stability threshold, it is determined that the host behavior has converged and is used as the host behavior map of the host for subsequent comparison baseline. In this embodiment, the stability threshold is preferably set to 0.05~0.10. For real-time traffic metadata, the corresponding host behavior graph is retrieved using the IP address of the host initiating the communication as an index to determine the degree of deviation between the current communication behavior and the host behavior graph. Extract the destination port of the current session and determine whether the destination port exists in the host's usual destination port distribution. If the destination port is a new port that has never appeared in the map, record it as a port deviation event. Extract the destination IP address of the current session, determine whether the destination IP address belongs to the communication peer IP set of this host, and if the destination IP address is a new peer node that has never appeared in the graph, record it as a peer deviation event; Extract the communication timestamp of the current session, map it to the corresponding hourly time period, and compare it with the historical frequency of the corresponding time period in the active time vector of the host. If the historical frequency of the corresponding time period is zero, that is, the host has never initiated communication in this time period, and the total load of the current session exceeds the preset abnormal traffic threshold, it is recorded as a time period deviation event. In this embodiment, the preset abnormal traffic threshold is preferably set to 2MB. When any two of the above deviation detections trigger deviation events simultaneously, it is determined that the host has undergone a behavioral gene mutation. The communication behavior records that triggered the deviation are extracted, including the source IP, destination IP, port and timestamp involved, and rule detection hit features are generated. This analysis summarizes all behavioral mutation records generated by intranet hosts within the current time window and performs horizontal correlation analysis from a network-wide perspective: All host behavioral mutation records are sorted in ascending order by communication timestamps to form a network-wide mutation event time sequence. Each mutation event includes the triggering host IP, deviation dimension type, and occurrence time. Using the host IPs in the mutation events as nodes and the communication quintuple relationships between adjacent mutation events as directed edges, a network-wide mutation propagation topology graph G is constructed. Specifically, if after a mutation event occurs on host A, within a preset propagation time threshold τ, host B, which has direct communication records with host A (verifiable from traffic metadata), also experiences a mutation event, then a directed propagation edge from A to B is established in G. The topological structure of the network-wide mutation propagation topology graph is analyzed. If it exhibits a tree-like or chain-like propagation structure starting from a single source node and spreading hierarchically to multiple surrounding nodes, it represents a typical intranet. Lateral movement characteristics indicate the presence of cross-host coordinated attack behavior. For lateral movement within an internal network (such as APT infiltration), a setting of 30 minutes to 2 hours is recommended, with 1 hour being preferred in this embodiment. All involved IP addresses and communication ports in the propagation link are extracted and compared with the local external threat intelligence database. The propagation topology, the list of nodes along the propagation path, and the intelligence database hit results are aggregated to generate malicious network characteristics. If a rule detection hit characteristic or a malicious network characteristic is generated, the threat matching result is configured as a known risk state, and this traffic data will be marked and removed in subsequent step S14. If no hit characteristic is generated, the threat matching result is configured as a risk-free state, and this traffic data will be processed as remaining traffic data in subsequent segmentation algorithms in step S14. The rule detection hit characteristics, malicious network characteristics, and threat matching results are structurally aggregated to generate potential threat data.
[0030] Step S2 includes the following sub-steps: Step S21: Extract the five-tuple and application layer attributes of traffic metadata, and perform deduplication and aggregation using the source IP address and destination IP address as indexes to obtain the set of host IPs of all active communication in the current intranet. For each host in the host IP set, extract the corresponding MAC address, operating system type and open service list to generate a set of online hosts in the intranet. Passive traffic analysis replaces active network scanning. The source and destination IP addresses of the five-tuple are extracted from the traffic metadata output by S12. After deduplication and aggregation, a set of IP addresses for all active communication hosts on the internal network is obtained. For example, in a real-world scenario, the following communication records appear in the traffic metadata within a certain time window, as shown in Table 1 below. Table 1 Communication Records: ; After deduplication and aggregation, the host IP set is {192.168.1.10, 192.168.1.20, 192.168.1.30, 192.168.1.40}, consisting of four active hosts. Subsequently, for each host, the operating system type is identified by TCP / IP protocol stack fingerprinting, and the open service list is identified by application layer protocol, generating a set of online hosts on the intranet.
[0031] The core of this embodiment lies in achieving zero-interference automatic discovery of intranet assets. Traditional asset management relies on manual entry or active scanning. The former has a lag, and the latter will generate additional traffic that interferes with business. By passively extracting asset information by analyzing existing traffic metadata, it not only ensures the real-time accuracy of the asset base, but also does not introduce any additional network burden. The set of online hosts in the intranet is the set of objects for profile modeling, and also the source of nodes for vulnerable node screening.
[0032] Step S22: Based on traffic metadata and traffic behavior characteristics, a dual-granularity fusion modeling method is used to establish an intranet behavior profile for the set of online hosts on the intranet. Step S221: Extract the session traffic status and communication timestamp of each host, and construct the vertical behavior sequence of each host on the time axis. The vertical behavior sequence is the change of port access frequency, uplink and downlink traffic ratio fluctuation and session duration distribution of the current host in different time windows. Through the communication frequency anomaly and session interaction relationship in the traffic behavior characteristics, mark the abnormal time sequence nodes in the vertical behavior sequence of the current host. In this embodiment, using a preset time window (1 hour in this embodiment) as the unit, time-series statistics are performed on the session traffic status and communication timestamp of each host in the traffic metadata output by S12 to construct a vertical behavior sequence.
[0033] Taking the intranet database server (192.168.1.40) as an example, the following behavioral data was obtained within a 7-day time window, as shown in Table 2 below; Table 2 Behavioral Data Statistics: ; The above statistical results constitute the baseline of the longitudinal behavior sequence of 192.168.1.40. Within the time window of 22:00 on a certain working day, the uplink traffic of the host was detected to suddenly increase to 8 times the normal average, and the uplink-to-downlink traffic ratio abnormally rose to 1:0.3, indicating a large amount of data uploading. This was consistent with the heartbeat back-connection pattern recorded in the communication frequency fluctuation characteristics. This time sequence node was marked as an abnormal time sequence node and recorded in the longitudinal behavior sequence. The longitudinal behavior sequence elevates host behavior from a single-point snapshot to a dynamic evolution trajectory on a time axis, enabling the system to identify periodic patterns and mutation modes of host behavior.
[0034] Step S222: Using all hosts in the set of online hosts in the intranet as nodes and the actual communication records between hosts as edges, extract the five-tuples from the traffic metadata to construct a network communication topology graph. Calculate the topology centrality index of each host in the network communication topology graph, and calculate the node value of each host in the intranet topology using the topology centrality index. The topology centrality index includes the number of communication connections, the frequency of communication relay nodes, and the average communication distance with other nodes. In this embodiment, all hosts in the set of online hosts in the intranet are used as nodes. The actual communication records between each host are extracted from the five-tuple of traffic metadata as directed edges to construct a network communication topology graph. Taking a simplified intranet with 5 hosts as an example, in the network communication topology graph, three types of topological centrality indicators are calculated for each host: the total number of edges that directly communicate between each host and other hosts is counted as the number of communication connections; the proportion of the number of shortest communication paths passing through the host to the total number of shortest paths in the entire network is counted as the communication relay node frequency; and the average of the shortest path lengths from the host to all other hosts in the entire network is calculated as the average communication distance.
[0035] The node value is calculated by weighted summation of the three types of topological centrality indices: ; in This is the normalized value of the number of communication connections. For communication relay node frequency, The average communication distance after min-max normalization. , , For the weighting coefficients, satisfying + + =1. In this embodiment, nodes with both high communication connection count and high relay frequency have the highest value and will be prioritized as simulation targets for mimicry decoys. Nodes with high communication connection count can directly affect the most surrounding hosts after being compromised. Nodes with high communication relay frequency are essential hubs for lateral movement within the intranet and have extremely high value to attackers. Nodes with short average communication distance can reach all network assets with the fewest hops and are ideal springboards for attackers to expand their gains. Three-dimensional fusion makes node value a precise quantification of node attractiveness from the attacker's perspective, serving as the basis for assigning node weights in the dynamic attack graph.
[0036] Step S223: The vertical behavior sequence of each host is fused with the node value to generate the intranet behavior profile of the current host. The intranet behavior profile includes host identity, open service list, vertical behavior sequence, node value and network topology location. The vertical behavioral sequence in the internal network behavior profile describes the past actions of the host, while the node value describes the importance of the host in the entire network. Only by combining the two can we show which host the attacker most wants to attack and how to detect anomalies. Step S23: Use a behavior dynamic mimicry engine to process the internal network behavior profile and generate a mimicry decoy.
[0037] Step S231: Traverse the intranet behavior profile, sort the node values in descending order, select the host with the highest node value as the target host, extract the open service list, port configuration and operating system type from the target host's intranet behavior profile as a mimicry feature set, construct a simulation instance in the virtualization container based on the mimicry feature set, and complete the static feature cloning. In this embodiment, the static cloning of the simulation instance not only copies the ports and services, but also includes the following details: the version number response of the MySQL service (5.7.32-log), the handshake fingerprint of the SSH protocol (SSH-2.0-OpenSSH_7.4), and the operating system fingerprint of the TCP / IP protocol stack (TTL value 64, TCP window size 29200). All of the above details are extracted from the application layer attributes of the target host's traffic metadata, so that the simulation instance presents a service fingerprint that is completely consistent with the real host under the detection of network scanning tools. Advanced attackers cannot distinguish the decoy from the real assets through conventional fingerprint identification methods. The simulation targets are selected based on descending node value to ensure that limited decoy resources are prioritized for deployment on high-value nodes that attackers are most likely to target, thus maintaining consistency with the optimal deployment strategy of the game model in terms of target selection logic.
[0038] Step S232: Continuously monitor the real-time traffic metadata of the target host and synchronize it to the corresponding simulation instance in real time; The dynamic synchronization mechanism is the key innovation of this embodiment. Traditional honeypots are cloned once during deployment and then remain fixed. However, the communication behavior of real business systems continues to evolve. Long-term lurking APT attackers can identify static honeypots by repeatedly probing and comparing the differences between the decoy's response and the real asset's response. This ensures that the mimicry decoy can maintain dynamic consistency with the real host after deployment and after the phantom matrix is reconstructed, supporting the continuous deception effectiveness throughout the entire trapping cycle.
[0039] Step S233: Based on the abnormal instructions and malicious payload distribution in the traffic behavior characteristics, a capture probe is pre-set in the simulation instance. The simulation instance and the corresponding mimicry feature set and capture probe configuration are encapsulated to generate a mimicry decoy.
[0040] The triggering conditions for the capture probe in this embodiment are designed as follows: When an instruction code matching the abnormal instruction characteristics appears in the operation instruction sequence of the access simulation instance, the capture probe is activated and the operation instruction sequence and context session data are fully recorded. When PE file header features or Base64 encoded executable file features are detected in the payload content of the simulation instance, the capture probe is activated to fully record the payload content and source quintuple. When the session interaction pattern of accessing the simulation instance deviates significantly from the baseline distribution of normal business access (e.g., the response latency variance is very small, exhibiting automated access characteristics), the capture probe is activated to record complete session data; The triggering conditions of the capture probe are directly derived from the traffic behavior characteristics. The simulation instance, mimicry feature set and capture probe configuration are encapsulated as a standardized mimicry decoy.
[0041] Step S3 includes the following sub-steps: Step S31: Construct a dynamic attack map using the network-wide communication topology map, potential threat data, and intranet behavior profiles. Step S32: Based on the dynamic attack graph, perform a probability assessment of the attack risk of each node, generate the node risk probability, and select a set of high-risk deployment nodes based on the node risk probability. Step S33: Based on the set of high-risk deployment nodes and the node risk probability, extract the attacker behavior type from the traffic behavior characteristics, construct the attacker type belief vector, introduce the business fluctuation disturbance factor, construct the belief-driven dynamic game model, and calculate the optimal decoy deployment strategy. Step S34: Based on the optimal decoy deployment strategy, deploy the mimicry decoy to the corresponding intranet location in the high-risk deployment node set to generate the decoy deployment topology.
[0042] Step S31: Extract the threat matching results from the potential threat data, filter out the hosts with known risk status from the set of online hosts in the intranet and mark them as vulnerable nodes. At the same time, traverse the intranet behavior profile, extract the open service list of each host, and mark the hosts with high-risk service exposure but no threat matching results as vulnerable nodes. Based on the overall network communication topology, and with vulnerable nodes as the starting point or intermediate stepping stone nodes of the attack path, the possible lateral movement paths of the attacker are deduced step by step along the directed communication edges in the overall network communication topology, generating a set of candidate attack paths. Using each host in the set of online hosts in the intranet as a node, and the lateral movement path in the set of candidate attack paths as a directed edge, the node value of each host in the intranet behavior profile is extracted as the node weight, and the high-risk port and vulnerable service information in the rule detection hit features are extracted as edge weights to generate a weighted directed dynamic attack graph. Step S32: Based on the dynamic attack graph, use Bayesian probabilistic inference to quantitatively assess the attack risk of each node: for each node in the dynamic attack graph... Calculate the node risk probability The recursive formula is: ; in, For nodes The set of predecessor nodes in the attack graph on that day For the predecessor node point to The directed edge weights are used to represent the probability of an attacker successfully moving laterally along the path. All nodes in the dynamic attack graph are arranged according to The nodes are sorted in descending order to filter out those whose risk probability exceeds the preset risk threshold θ, a node risk probability distribution is generated, and the filtered nodes are aggregated into a set of high-risk deployment nodes. For the set of predecessor nodes in a dynamic attack graph For nodes that are empty, initialize them according to their marked state in step S31: If the node is marked as a vulnerable node in step S31, then let This is equal to the preset initial risk value of the vulnerable node; in this embodiment, the value is 0.8. If the node is not marked as a vulnerable node, but its corresponding open service list contains high-risk service ports, then... This is equal to the preset initial risk value for high-risk services; in this embodiment, the value is set to 0.5. The remaining nodes It is equal to the preset baseline risk value, which is 0.1 in this embodiment; After initialization, the above recursive formula is applied sequentially from smallest to largest according to the shortest path length from each node to the vulnerable node in the dynamic attack graph. In this embodiment, a vulnerable node refers to a host that has been detected by the dual-mode matching engine as having actual attack traces, or a host with known threat matching results. This status indicates that the attacker has conducted reconnaissance, penetration, or exploitation activities on the node, and the probability of the node being compromised is significantly higher than that of a host with exposed services. High-risk service nodes are only marked because they have open high-risk ports or vulnerable services. Although there is potential for attack, no actual attack activity has been observed. Therefore, in the initialization of the starting node without a preceding propagation path, this embodiment sets a higher initial risk probability of 0.8 for vulnerable nodes, reflecting the real risk of being actually contacted by the attacker. The risk probability of high-risk service nodes is 0.5, and the remaining baseline nodes are the lowest at 0.1. This hierarchical initialization enables subsequent events to more accurately reflect the real attack and defense situation and avoids the problem of underestimating the risk of ignoring attack traces.
[0043] When deploying a mimicry decoy, an idle IP address in the current internal subnet is assigned as the decoy IP, or a flow table is issued by a software-defined network controller to statically route only traffic with attack characteristics or traffic from a specific source IP to the mimicry decoy's MAC address, thus avoiding IP / MAC conflicts with real high-risk hosts.
[0044] Step S33: Extract abnormal command sequences, API unauthorized call patterns, malicious payload distribution, and session interaction relationships from traffic behavior features to construct a feature vector for each attack behavior; Unsupervised clustering of the above feature vectors yields K types of attacker behavior. Each type corresponds to an attacker's decision-making pattern when moving laterally within the internal network, such as rapid scanning, slow-speed stealth, and targeted penetration; each type This corresponds to the prior payoff distribution of an attacker targeting each node in this mode. ; The defender maintains a belief vector about the attacker's type. ,in ( The attacker is of type [type]. The probability satisfies Where K is the total number of attacker behavior types, k is the type index number, and the attacker type belief vector is initialized based on the historical threat records of each vulnerable node in the node risk probability. Initialize to a uniform distribution; whenever the mimicry decoy's capture probe records a new attack interaction, extract the feature observation value 'o' of that interaction, and use Bayesian filtering to update the belief vector online: ; in, For attacker type The likelihood probability of observing behavioral feature o under the given conditions is derived from the prior benefit distributions of each type. It is calculated that as attack interaction data continues to accumulate, the belief vector b will continuously converge, and the defender's judgment of the attacker's true intentions will become increasingly accurate. From the session traffic status of traffic metadata, the uplink and downlink traffic ratio fluctuations of each host node within the current time window are statistically analyzed in real time. Combined with the abnormal time-series nodes marked in the vertical behavior sequence, the overall business fluctuation disturbance factor of the current intranet is calculated. : ; in For nodes The standard deviation of the flow rate in the current time window. The historical average traffic value corresponding to its longitudinal behavior sequence. To keep the historical flow average constant, and prevent it from being zero, let The business fluctuation level of each node is continuously mapped to [0,1). The larger the value, the more drastic the fluctuations in internal network services, and the easier it is for attackers to mix malicious traffic with normal business traffic. The attacker's mixed expected payoff is calculated based on the belief vector b; the defender at the node The probability of deploying a mimicry decoy is The attacker attacks the node driven by belief vector b. The weighted expected return is: ; The defender's belief-driven expected payoff function is: ; in Based on Calculate the probability of a mixed attack by the attacker. The defensive benefits when a mimicry decoy successfully lures a prey. For nodes Defensive losses during the attack The resource cost of deploying mimicry decoys, For business fluctuation disturbance factors False alarm penalty for the driven node; Risk probability of each node in the node risk probability distribution As the initial input to the belief state space, the future returns of long-term APT penetration are discounted using a preset discount factor δ. The AD-DSG algorithm is used for dynamic equilibrium iterative solution. In each iteration, the attacker's mixed return is updated based on the current belief vector b. Simultaneously, the node penalty term is dynamically adjusted based on business fluctuation disturbance factors until the change in revenue falls below the convergence threshold. (Preferably set to 0.01), the optimal decoy deployment strategy is obtained. In this embodiment, the discount factor is preferably set to 0.95. Based on the optimal decoy deployment strategy, the mimicry decoys are scheduled according to the deployment priority of each node; Specifically, based on the deployment probability of each node in the optimal decoy deployment strategy. The instances of the mimicry decoys, which encapsulate the corresponding mimicry feature set and capture probe configuration, are first bound to the highest priority target node in the high-risk deployment node set. After the mimicry decoys are deployed to the target node, their service response behavior is kept in dynamic consistency with the surrounding real assets based on the real-time traffic metadata of the target host that is continuously synchronized in step S232. The location information of the deployed mimicry decoy nodes, their corresponding mimicry feature sets, and their topology binding relationships in the intranet are recorded in a structured manner to generate the decoy deployment topology.
[0045] Step S4 includes the following sub-steps: Step S41: Based on the decoy deployment topology, continuously monitor the access traffic of the mimic decoy, extract attack indicators and record them as network attack behaviors; In this embodiment, based on the decoy deployment topology, the pre-installed capture probes on all mimicry decoy nodes in the topology are continuously activated for listening. Since the mimicry decoy simulates high-value real assets in the intranet, the historical communication records of normal business hosts in the intranet, i.e., the set of communication peer IPs in the intranet behavior profile, do not include access to the decoy nodes. Therefore, any access behavior that triggers the mimicry decoy has a very high probability of being malicious in the prior art, and no additional false alarm filtering mechanism is needed. Taking a specific scenario as an example, the host controlled by the attacker (192.168.1.60) initiates a MySQL connection request to the mimicry decoy deployed on 192.168.1.40. The capture probe captures the complete data of the session in real time, extracts attack indicators, including the attack source IP, attack timestamp, the identifier of the triggered mimicry decoy node, the operation command sequence and payload content, and integrates the above attack indicators into the network attack behavior, which is stored in the database in real time.
[0046] Step S42: Infer intent from attack indicators in network attack behavior to generate attack tactical intent; Step S42 specifically includes: Extract the sequence of operation instructions and payload content from network attack behavior, preset the proportion threshold, including the first proportion threshold, the second proportion threshold and the third proportion threshold, count the proportion of the occurrence frequency of each type of instruction in the operation instruction sequence to the sum of the occurrence frequencies of all instructions, and identify the attack stage, including the first stage, the second stage and the third stage. When the sum of the frequencies of port scanning commands and service probe commands exceeds the first ratio threshold, and the payload content does not contain vulnerability exploitation feature fields, the current attack stage is determined to be the first stage. When the sum of the frequencies of exploit commands, lateral scanning commands, and credential retrieval commands exceeds the second proportion threshold, and the payload contains exploit feature fields that match the open service list of the mimic decoy, the current attack phase is determined to be the second phase. When the frequency of data outbound transmission commands exceeds the third proportion threshold, and the length of a single payload in the payload content exceeds the preset payload length threshold, the current attack stage is determined to be the third stage. The attack phase determination results are structurally aggregated with the host nodes in the internal network behavior profile that have communication records with the attack source IP to generate attack tactical intent.
[0047] After initially gaining access to the internal network, the attacker initiated a series of probing operations on the decoy node. The sequence of operation commands captured by the probes was as follows: port scan commands appeared 142 times, accounting for 58.7%; service probe commands appeared 31 times, accounting for 12.8%; and other commands appeared 69 times, accounting for 28.5%. The combined frequency of port scanning commands and service probe commands accounts for 71.5%, which is greater than the preset first proportion threshold (preferably set to 65% in this embodiment), and the payload content does not contain any vulnerability exploitation feature fields. Therefore, the current attack stage is determined to be the first stage, namely the reconnaissance and detection stage.
[0048] By combining the internal network behavior profile, host nodes that have communication records with the attack source IP are extracted, and it is inferred that the attacker may launch a probe against the above two hosts next. The inference result is aggregated with the attack stage judgment result to generate the attack tactical intent. As the attacker completes initial reconnaissance and transitions to active penetration, the sequence of operational instructions changes significantly, entering the second stage of judgment. The sequence of operational instructions becomes: exploit commands appear 89 times, accounting for 41.2%; lateral scanning commands appear 47 times, accounting for 21.8%; credential attempt commands appear 38 times, accounting for 17.6%; and other commands appear 42 times, accounting for 19.4%. The sum of the frequencies of exploit commands, lateral scanning commands, and credential attempt commands accounts for 80.6%, which is greater than the preset second proportion threshold (preferably set to 70% in this embodiment). Furthermore, the payload content contains SQL injection vulnerability exploitation feature fields (such as UNION SELECT and INTO OUTFILE feature strings) that match the MySQL service in the mimicry decoy's open service list. The current attack stage is determined to be the second stage, namely the penetration and diffusion stage; the attack tactical intent is updated.
[0049] After completing lateral movement, the attackers entered the data theft phase. Data outbound transmission commands appeared 178 times, accounting for 73.6% of the total; other commands appeared 64 times, accounting for 26.4%. The proportion of data outbound transmission instructions is 73.6%, which is greater than the preset third proportion threshold (preferably set to 60% in this embodiment), and the length of a single payload in the payload content reaches 48MB, which exceeds the preset payload length threshold (preferably set to 10MB). Therefore, the current attack stage is determined to be the third stage, namely the data theft stage.
[0050] The three-stage framework for determining attack tactical intent covers the entire lifecycle of an APT attack. Through a combined determination mechanism of command frequency statistics and payload characteristic verification, it avoids the risk of false positives from single-dimensional judgments. Taking the second stage of determination as an example, relying solely on command frequency might misjudge many normal database maintenance operations as penetration activities. Introducing vulnerability exploitation feature field matching within the payload content as a second layer of verification minimizes the risk of false positives.
[0051] Step S43: Based on the deception defense information entropy model, continuously calculate the entropy value of the current trapping scenario, evaluate the deception effectiveness of the mimicry decoy, and trigger the reconstruction of the phantom matrix when the entropy value is lower than the preset safety threshold. If there are M decoy nodes in the system, the maximum information entropy is It is recommended to set the preset security threshold as a percentage of the maximum entropy value, for example... When the probability of attacker access is highly concentrated, causing the system entropy to drop to half of the ideal random state, a reconstruction is triggered. Before calculating the deception defense information entropy, a Laplace smoothing initialization of the access probability is introduced. Using the probability distribution of each mimicry decoy node in the current decoy deployment topology being accessed by the attacker as input, the information entropy value of the current trapping scenario is continuously calculated. : ; in The probability that an attacker chooses to access the j-th mimicry decoy node, where M is the total number of mimicry decoys in the current decoy deployment topology; When the probability of an attacker accessing each mimicry decoy node tends to be uniform, When the probability of access approaches its maximum value, it indicates that the attacker has not yet figured out the trapping scenario; when the access probability is highly concentrated on a few nodes, The continuous decline indicates that the attacker has accumulated sufficient environmental knowledge and is at risk of recognizing the decoy. The entropy-driven proactive reconstruction mechanism enables the system to trigger a scenario update before the deception effectiveness completely fails.
[0052] Step S44: Dynamically construct a phantom matrix based on the attack tactical intent and the internal network behavior profile; This embodiment infers, based on the attacker's tactical intent, that the attacker might initiate a lateral movement towards the file server, which has historically had extensive data exchange with the database server and is highly correlated with Node-40 in the internal network behavior profile. Asset characteristics of the file server with the second-highest node value are extracted from the internal network behavior profile to generate a new mimicry feature set. A new simulation instance is constructed in a virtualization container and injected into key locations in the attacker's predicted exploration path. Furthermore, according to the continuous synchronization mechanism in step S232, the real-time traffic metadata of 192.168.1.50 is synchronized to the new simulation instance. Continuous access by attackers The Node-40 with the highest contribution value, P(s1)=0.68, has the most severe deception effectiveness decay. Based on the target host's real-time traffic metadata that is continuously synchronized in step S232, its service configuration is reconstructed from MySQL (port 3306) to PostgreSQL (port 5432), the operating system fingerprint is updated from Linux CentOS 7 to Linux Ubuntu 20.04, and the response latency characteristics are re-randomized. The reconstructed Node-40 presents itself to the attacker as a completely new database server, rendering the attacker's accumulated knowledge invalid.
[0053] The newly injected file server simulation instance is integrated with the reconstructed Node-40 and the remaining three original decoy nodes. The decoy deployment topology is updated to form a phantom matrix covering the attacker's current exploration path. The reconstructed phantom matrix node distribution exposes the attacker to a completely new internal network environment, rendering most of their accumulated reconnaissance information invalid and forcing them to conduct reconnaissance and probing again. It then recovers to near its maximum entropy value, and the deception effectiveness is restored.
[0054] Step S45: Input all communication traffic corresponding to the network attack behavior into the phantom matrix to complete the isolation of the network attack behavior.
[0055] In this embodiment, all subsequent communication traffic corresponding to the attack source IP of the network attack is transparently proxied, diverting the communication traffic to the phantom matrix. The implementation of the transparent proxy is completely imperceptible to the attacker. All network connection requests initiated by the attacker on 192.168.1.60 are silently redirected at the network layer to the corresponding simulation instance in the phantom matrix. All responses received by the attacker come from the phantom matrix, not the real internal network assets. Taking the attacker's attempt to initiate a lateral movement from Node-40 (which has been reconstructed as a PostgreSQL service) to the real internal network file server (192.168.1.50) as an example, the connection request is transparently redirected by the proxy to the newly injected file server simulation instance in the phantom matrix. All of the attacker's lateral movement operations are consumed in the fake environment and have no substantial impact on the real internal network assets (192.168.1.50). The attacker's continuous operations in the phantom matrix are still recorded in real time by the capture probes of each node and updated in the network attack behavior record.
[0056] Step S5 includes the following sub-steps: Step S51: Integrate network attack behavior, decoy deployment topology, and traffic metadata to reconstruct the attacker's complete operation sequence and generate attack movement trajectory; Step S52: Retrieve attack movement trajectory and potential threat data, and output security event records; Step S53: Feed back the attack tactical intent to step S13 and update the host behavior graph; Step S54: Summarize potential threat data, attack movement trajectories, and attack tactical intentions to generate a cybersecurity detection report.
[0057] This embodiment integrates three data sources—network attack behavior, decoy deployment topology, and traffic metadata—in step S51 for time-series fusion and reconstruction, solving the problem of fragmented attack event records and the inability to form a complete attack chain in traditional security tools. This ensures that the entire path of the attacker, from initial intrusion, lateral movement, decoy triggering, to isolation by the phantom matrix, is completely preserved. Based on this, step S52 supports arbitrary combinations of retrieval based on time range, attack source IP, target service type, and attack stage annotations, transforming massive amounts of security event data into a structured knowledge base that can be searched on demand, significantly reducing the time cost of security event tracing and forensics. Step S53… Furthermore, the novel attack methods captured during the trapping process are automatically converted into updated samples of the host behavior graph, which are fed back to the dual-mode matching engine in step S13. This allows the system to continuously respond to the evolution of attacker methods without manual intervention, achieving autonomous growth in defense capabilities. Finally, step S54 integrates potential threat data, attack movement trajectories, and attack tactical intentions into a standardized network security detection report containing four parts: asset risk overview, attack event statistics, attack path reconstruction, and security remediation suggestions. This achieves a complete link from underlying traffic characteristics to management-level security decisions, providing complete data support for internal network security remediation and the next round of defense strategy optimization.
[0058] Example 2, refer to Figure 2 This paper provides an intelligent network security detection system based on traffic analysis, including: The traffic awareness module is used to parse the obtained traffic metadata, analyze potential threat data, and extract traffic behavior features based on a segmentation algorithm. The profile generation module is used to identify online hosts on the intranet through traffic metadata and traffic behavior characteristics, and to build intranet behavior profiles and dynamically generate mimicry decoys; The attack game module is used to generate dynamic attack graphs based on potential threat data and internal network behavior profiles, and to deploy mimicry decoys to high-risk nodes in the internal network. The decoy isolation module is used to detect access traffic directed to the mimicry decoy, obtain network attack behavior, construct a phantom matrix, and isolate network attack behavior; The attack tracing module is used to save the movement trajectory of network attack behavior and generate network security detection reports.
[0059] In this embodiment, the traffic awareness module can capture bidirectional traffic on the intranet in real time and parse it to obtain traffic metadata. Simultaneously, it uses a dual-mode matching engine to analyze potential threat data and extracts traffic behavior characteristics through a segmentation algorithm, ensuring comprehensive awareness of known threats and unknown protocol traffic. The profile generation module automatically identifies online hosts on the intranet based on traffic metadata and traffic behavior characteristics, establishes intranet behavior profiles, and dynamically generates mimicry decoys, achieving automatic asset inventory management and high-fidelity decoy construction without manual scanning or active probing, avoiding additional network burden. The attack game theory module, based on potential threat data and intranet behavior profiles... The system generates a dynamic attack graph and uses a belief-driven dynamic game model to optimize the deployment of mimicry decoys to high-risk nodes on the internal network, improving the utilization efficiency and targeting of limited decoy resources. The decoy isolation module continuously monitors access traffic directed to the mimicry decoys, obtains network attack behavior, and constructs a phantom matrix based on attack tactical intent and deception defense information entropy to isolate attack traffic within the phantom matrix, effectively protecting real business assets. The attack tracing module saves the movement trajectory of network attack behavior and generates a network security detection report by combining potential threat data, providing complete data support for security incident tracing, compliance auditing, and hidden danger rectification.
[0060] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.
[0061] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.
[0062] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.
Claims
1. A traffic analysis based network security intelligent detection method, characterized in that, Includes the following steps: Step S1: Collect and parse the bidirectional network traffic data of the intranet to obtain traffic metadata. Analyze potential threat data through a dual-mode matching engine and extract traffic behavior features based on a segmentation algorithm. Step S2: Based on the traffic metadata and traffic behavior characteristics, automatically identify online hosts on the intranet, establish intranet behavior profiles for each host, and dynamically generate mimicry decoys; Step S3: Generate a dynamic attack graph based on potential threat data and intranet behavior profiles, and deploy the mimicry decoy to high-risk nodes in the intranet using a game theory model; Step S4: Continuously detect access traffic pointing to the mimic decoy, obtain network attack behavior, and use dynamic trapping technology to construct a phantom matrix and isolate the network attack behavior; Step S5: Save the movement trajectory of the network attack behavior and generate a network security detection report; Step S2 includes the following sub-steps: Step S21: Extract the five-tuple and application layer attributes of the traffic metadata, and perform deduplication and aggregation using the source IP address and destination IP address as indexes to obtain the set of host IPs of all active communication in the current intranet. For each host in the set of host IPs, extract the corresponding MAC address, operating system type and open service list to generate a set of online hosts in the intranet. Step S22: Based on the traffic metadata and traffic behavior characteristics, a dual-granularity fusion modeling method is used to establish an intranet behavior profile for the set of intranet online hosts; Step S221: Extract the session traffic status and communication timestamp of each host, and construct the vertical behavior sequence of each host on the time axis. The vertical behavior sequence is the port access frequency change, uplink and downlink traffic ratio fluctuation and session duration distribution of the current host in different time windows. Through the communication frequency anomaly and session interaction relationship in the traffic behavior characteristics, mark the abnormal time sequence nodes in the vertical behavior sequence of the current host. Step S222: Using all hosts in the set of online hosts in the intranet as nodes and the actual communication records between hosts as edges, extract the five-tuples from the traffic metadata to construct a full network communication topology graph. Calculate the topology centrality index of each host in the full network communication topology graph, and calculate the node value of each host in the intranet topology using the topology centrality index. The topology centrality index includes the number of communication connections, the frequency of communication relay nodes, and the average communication distance with other nodes. Step S223: The vertical behavior sequence of each host is fused with the node value to generate the intranet behavior profile of the current host. The intranet behavior profile includes host identity, open service list, vertical behavior sequence, node value and network topology location. Step S23: The behavior profile of the intranet is processed using a behavior dynamic mimicry engine to generate a mimicry decoy; Step S231: Traverse the intranet behavior profile, sort the node values in descending order, select the host with the highest node value as the target host, extract the open service list, port configuration and operating system type from the intranet behavior profile of the target host as a mimicry feature set, construct a simulation instance in the virtualization container based on the mimicry feature set, and complete the static feature cloning. Step S232: Continuously monitor the real-time traffic metadata of the target host and synchronize it to the corresponding simulation instance in real time; Step S233: Based on the abnormal instructions and malicious payload distribution in the traffic behavior characteristics, a capture probe is pre-set in the simulation instance, and the simulation instance and the corresponding mimicry feature set and capture probe configuration are encapsulated to generate a mimicry decoy; Step S3 includes the following sub-steps: Step S31: Construct a dynamic attack map using the potential threat data and intranet behavior profile through the network-wide communication topology map; Step S32: Based on the dynamic attack graph, perform a probability assessment on the attack risk of each node, generate the node risk probability, and select a set of high-risk deployment nodes based on the node risk probability. Step S33: Based on the set of high-risk deployment nodes and node risk probabilities, extract attacker behavior types from traffic behavior characteristics, construct attacker type belief vectors, introduce business fluctuation disturbance factors, construct belief-driven dynamic game model, and calculate the optimal decoy deployment strategy. Step S34: According to the optimal decoy deployment strategy, deploy the mimicry decoy to the corresponding intranet location in the high-risk deployment node set to generate the decoy deployment topology; The processing logic of the dual-mode matching engine is as follows: Extract the five-tuple and corresponding timestamp of the traffic metadata, and construct a host behavior graph based on a sliding time window with each host in the intranet as a node and communication behavior as an edge. The host behavior graph includes port distribution, set of communication peer IPs and active time periods. Obtain the metadata of the currently arriving traffic and extract the current communication behavior. Traverse the host behavior graph. When no sample in the host behavior graph has the same communication pattern as the current communication behavior, it is determined to be a behavioral gene mutation, and a rule is generated to detect the hit feature. Extract communication behaviors corresponding to behavioral gene mutations from multiple hosts, perform correlation analysis, and generate malicious network characteristics; Based on the rules, the system detects hit features and malicious network features, generates threat matching results, and packages them to generate potential threat data.
2. The network security intelligent detection method based on traffic analysis of claim 1, wherein, Step S1 includes the following sub-steps: Step S11: Capture host-to-host traffic data as first data, capture host-to-gateway traffic data as second data, and use the first data and the second data as intranet bidirectional network traffic data. Step S12: Decompose and decode the intranet bidirectional network traffic data, extract the header information and session flow status of each layer and perform standardization processing to obtain traffic metadata, which includes a five-tuple, session traffic status, application layer attributes and protocol parsing status identifier. Step S13: Analyze the traffic metadata using a dual-mode matching engine to generate potential threat data, which includes threat matching results, rule detection hit characteristics, and malicious network characteristics. Step S14: From the intranet bidirectional network traffic data, remove traffic data with known risks by matching the protocol parsing status of the traffic metadata with the potential threat data, filter out the remaining traffic data, and call the segmentation algorithm to cut the remaining traffic data to obtain the corresponding functional fields; Step S15: By summarizing and extracting the functional fields and traffic metadata, traffic behavior characteristics are obtained. The traffic behavior characteristics include abnormal commands, API unauthorized calls, malicious payload distribution, communication frequency fluctuations, and session interaction relationships.
3. The network security intelligent detection method based on traffic analysis of claim 2, wherein, Step S4 includes the following sub-steps: Step S41: Based on the decoy deployment topology, continuously monitor the access traffic of the mimic decoy, extract attack indicators and record them as network attack behaviors; Step S42: Infer intent from the attack indicators in the network attack behavior to generate attack tactical intent; Step S43: Based on the deception defense information entropy model, continuously calculate the entropy value of the current trapping scenario, evaluate the deception effectiveness of the mimicry decoy, and trigger the reconstruction of the phantom matrix when the entropy value is lower than the preset safety threshold. Step S44: Based on the attack tactical intent, dynamically construct a phantom matrix in conjunction with the intranet behavior profile; Step S45: Input all communication traffic corresponding to the network attack behavior into the phantom matrix to complete the isolation of the network attack behavior.
4. The network security intelligent detection method based on traffic analysis of claim 3, wherein, Step S42 specifically includes: Extract the operation instruction sequence and payload content from the network attack behavior, preset the proportion threshold, the proportion threshold includes a first proportion threshold, a second proportion threshold and a third proportion threshold, count the proportion of the occurrence frequency of each type of instruction in the operation instruction sequence to the sum of the occurrence frequencies of all instructions, and identify the attack stage, the attack stage includes a first stage, a second stage and a third stage. When the sum of the frequencies of port scanning commands and service probe commands is greater than a first ratio threshold, and the payload content does not contain vulnerability exploitation feature fields, the current attack stage is determined to be the first stage. When the sum of the frequencies of vulnerability exploitation commands, lateral scanning commands, and credential retrieval commands exceeds a second threshold, and the payload content contains vulnerability exploitation feature fields that match the open service list of the mimic decoy, the current attack phase is determined to be the second phase. When the frequency of data outbound transmission commands exceeds the third proportion threshold, and the length of a single payload in the payload content exceeds the preset payload length threshold, the current attack stage is determined to be the third stage. The attack phase determination results are structurally aggregated with the host nodes in the internal network behavior profile that have communication records with the attack source IP to generate attack tactical intent.
5. The network security intelligent detection method based on traffic analysis of claim 4, wherein, Step S5 includes the following sub-steps: Step S51: Integrate the network attack behavior, the decoy deployment topology, and the traffic metadata to reconstruct the attacker's complete operation sequence and generate an attack movement trajectory; Step S52: Retrieve the attack movement trajectory and the potential threat data, and output the security event record; Step S53: Feed back the attack tactical intent to step S13 and update the host behavior graph; Step S54: Summarize the potential threat data, attack movement trajectory, and attack tactical intent to generate a network security detection report.
6. A network security intelligent detection system based on traffic analysis, which is applied to the network security intelligent detection method based on traffic analysis as claimed in any one of claims 1-5, characterized in that, It includes a traffic awareness module, a profile generation module, an attack game theory module, a trapping and isolation module, and an attack tracing module; The traffic awareness module is used to parse the traffic metadata, analyze potential threat data, and extract traffic behavior features based on a segmentation algorithm. The profile generation module is used to identify online hosts on the intranet through traffic metadata and traffic behavior characteristics, and to establish intranet behavior profiles and dynamically generate mimicry decoys. The attack game module is used to generate dynamic attack graphs based on potential threat data and internal network behavior profiles, and to deploy mimicry decoys to high-risk nodes in the internal network. The trapping and isolation module is used to detect access traffic directed to the mimicry decoy, obtain network attack behavior, construct a phantom matrix, and isolate network attack behavior. The attack tracing module is used to save the movement trajectory of the network attack behavior and generate a network security detection report.
Citation Information
Patent Citations
Method, device and equipment for processing auditing task of network security compliance file
CN122394876A
Honeypot bait distribution method and device, storage medium and electronic equipment
CN113037777A
