A cloud-edge collaborative method, device, and system for cross-domain threat perception of power terminals.
By employing a cloud-edge collaborative approach to cross-domain threat perception in power terminals, and utilizing the identification and status features collected by edge terminals to construct a full-element perception vector, the problem of identifying and tracing deep-seated security risks in cross-regional and cross-domain attacks in power Internet of Things systems is solved, achieving efficient threat identification and immediate response.
Patent Information
- Application Number
- CN202611239663.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-17
AI Technical Summary
Existing power Internet of Things (IoT) systems struggle to accurately identify and trace deep-seated security risks in cross-regional and cross-domain attacks, and insufficient cloud-edge computing collaboration leads to high response latency, making it impossible to instantly block high-risk attacks.
A cloud-edge collaborative approach for cross-domain threat perception of power terminals is adopted. By collecting business terminal identifiers and operational status characteristics through edge terminals, an identifier mapping relationship is established and associated with the cloud master station to construct a full-element perception vector. Threat analysis is performed using a temporal topology fusion constraint network to achieve accurate identification and tracing of cross-domain threats.
It improves the accuracy of cross-domain attack identification and tracing capabilities, enables millisecond-level local handling of high-risk attacks, eliminates time window vulnerabilities in security response, and improves the security monitoring efficiency of power terminals.
Smart Images

Figure CN122824498B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power Internet of Things (IoT) security monitoring technology, specifically to a cloud-edge collaborative method, device, and system for cross-domain threat perception of power terminals. Background Technology
[0002] With the in-depth construction of new power systems, a massive number of distributed power generation terminals, distribution automation terminals, and load-side sensing terminals are widely connected to the power grid, forming a typical cloud-network-edge-device business architecture for the power Internet of Things (IoT). Because these terminals are mostly located in geographically remote, physically open production sites, their inherent security protection is weak and they lack real-time security monitoring methods. This causes the security boundary of the power system to extend infinitely to the physical end, expanding the exposure surface for network attacks. Attackers can easily use controlled edge terminals as springboards for lateral movement through physical contact or wireless penetration methods, and then use cloud-edge communication tunnels to breach the main station's defenses, posing a serious penetration threat to core infrastructure such as the power grid dispatch center.
[0003] Currently, in the cloud-network-edge-device architecture of power IoT systems, edge security equipment such as IPS (Intrusion Prevention System) and WAF (Web Application Firewall) are typically deployed at cloud-network nodes to monitor abnormal terminal traffic. At edge-device nodes, security monitoring probes are mainly deployed and run to monitor operational risks such as malicious processes and high-risk ports on the terminals. Under this monitoring model, security analysis is singular and fragmented. When a terminal is subjected to protocol instruction tampering or covert channel attacks, the master station can often only capture surface-level traffic fluctuations and cannot obtain the root cause of the anomaly. This makes it difficult for existing systems to accurately identify and trace deeper security risks in complex cross-regional and cross-domain attacks. Summary of the Invention
[0004] To overcome the problem of low accuracy in identifying and tracing deep-seated security risks in complex cross-regional and cross-domain attacks by existing security monitoring systems, this invention provides a cloud-edge collaborative method, device, and system for cross-domain threat perception of power terminals.
[0005] In a first aspect, the present invention provides a cloud-edge collaborative method for cross-domain threat perception of power terminals, applied to a cloud-based main station, comprising: The system receives the identifier mapping relationship, the operating status characteristics, and the edge threat score reported by the edge terminal; the identifier mapping relationship is used to indicate the association relationship between the edge terminal and each service terminal accessing the edge terminal. If it is determined that the edge threat score of an edge terminal is within the medium risk threshold range, the monitored service traffic is analyzed to extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic. Based on the identification mapping relationship, the service traffic characteristics of each service terminal are associated with the operating status characteristics of the edge terminal associated with the service terminal identification to construct a perception vector; The perception vector is input into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal.
[0006] Secondly, the present invention provides a cloud-based main station, comprising: The main station traffic mirroring probe is used to receive the identifier mapping relationship, the operating status characteristics, and the edge threat score reported by the edge terminal. The identifier mapping relationship is used to indicate the association relationship between the edge terminal and each service terminal accessing the edge terminal. It is also used to analyze the monitored service traffic and extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic if it is determined that the edge threat score of an edge terminal is within the medium risk threshold range. Based on the identifier mapping relationship, the service traffic characteristics of each service terminal are associated with the operating status characteristics of the edge terminal associated with the service terminal identifier to construct a perception vector. The terminal anomaly risk deep analysis engine is used to input the perception vector into a temporal topology fusion constraint network for threat analysis, and obtain the threat identification result of the edge terminal.
[0007] Thirdly, this invention provides a cloud-edge collaborative method for cross-domain threat perception of power terminals, applied to edge terminals, including: The monitored service traffic is analyzed to extract the service terminal identifiers of each service terminal accessing the edge terminal, and the service terminal identifiers are associated with the edge terminal to generate an identifier mapping relationship; The edge terminal collects its own operational status characteristics and calculates its edge threat score based on the degree of deviation of the operational status characteristics from the baseline. The identification mapping relationship, the operational status characteristics, and the edge threat score are reported to the cloud master station, so that the cloud master station can analyze the monitored service traffic, extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic, and associate the service traffic characteristics of each service terminal with the operational status characteristics of the edge terminal associated with the service terminal identifier based on the identification mapping relationship to form a perception vector, and input the perception vector into the temporal topology fusion constraint network for threat analysis.
[0008] Fourthly, the present invention provides an edge terminal, including a terminal security monitoring probe, the terminal security monitoring probe comprising: The mapping generation unit is used to parse the monitored service traffic, extract the service terminal identifiers of each service terminal accessing the edge terminal, and associate the service terminal identifiers with the edge terminal to generate an identifier mapping relationship. The threat calculation unit is used to collect its own operational status characteristics and calculate the edge threat score of the edge terminal based on the degree of deviation of the operational status characteristics from the baseline. The reporting unit is used to report the identifier mapping relationship, the operating status characteristics, and the edge threat score to the cloud master station, so that the cloud master station can analyze the monitored service traffic, extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic, and associate the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identifier based on the identifier mapping relationship to form a perception vector, and input the perception vector into the temporal topology fusion constraint network for threat analysis.
[0009] Fifthly, the present invention provides a cloud-edge collaborative method for cross-domain threat perception of power terminals, comprising: The edge terminal parses the monitored service traffic, extracts the service terminal identifiers of each service terminal accessing the edge terminal, and associates the service terminal identifiers with the edge terminal to generate an identifier mapping relationship; The edge terminal collects its own operating status characteristics and calculates the edge threat score based on the degree of deviation of the operating status characteristics from the baseline; the identification mapping relationship, the operating status characteristics, and the edge threat score are reported to the cloud master station. When the cloud master station determines that the edge threat score of an edge terminal is within the medium risk threshold range, it analyzes the monitored business traffic and extracts the business traffic characteristics of each business terminal and the business terminal identifier carried in the business traffic. Based on the identifier mapping relationship, the cloud master station associates the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identifier to construct a perception vector; and inputs the perception vector into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal.
[0010] In a sixth aspect, the present invention provides a cloud-edge collaborative power terminal cross-domain threat perception system, comprising: the aforementioned cloud master station and / or the aforementioned edge terminal.
[0011] In a seventh aspect, the present invention provides a readable storage medium having an executable program stored thereon, wherein when the executable program is executed, it implements the method described in any one of the foregoing descriptions.
[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a cloud-edge collaborative method and system for cross-domain threat perception of power terminals. It receives the identification mapping relationship between service terminals and edge terminals reported by edge terminals. The cloud master station can use this mapping relationship as a bridge to accurately correlate the operating status characteristics from edge terminals with the service traffic characteristics from the master station, constructing a dual-domain perception vector covering both internal terminal operation anomalies and external network traffic anomalies. This perception vector is then input into a temporal topology fusion constraint network for threat analysis, yielding the threat identification result for the edge terminal. This allows for the accurate reproduction of the lateral movement path of attacks in cross-domain data fusion analysis, improving the accuracy of cross-domain attack identification and the ability to deeply trace attack paths.
[0013] By decentralizing the calculation of edge threat scores to edge terminals, the cloud main station only calls the computationally expensive time-series topology fusion network for in-depth analysis of edge terminals located in the medium-risk threshold range. This achieves hierarchical collaborative processing of coarse edge scoring and fine cloud network analysis, balancing immediate attack response and accurate identification of terminal security risks in scenarios with massive terminal access. Attached Figure Description
[0014] Figure 1 This is a flowchart illustrating a cloud-edge collaborative cross-domain threat perception method for power terminals according to the present invention. Figure 2 This is a schematic diagram of a temporal topology fusion constraint network structure and processing procedure, as exemplified by the present invention. Figure 3 This is a schematic diagram of a cloud-edge collaborative service terminal threat perception method according to an example of the present invention; Figure 4 This is a schematic diagram of a cloud-edge collaborative service terminal threat perception system, which is an example of the present invention. Detailed Implementation
[0015] Currently, in the cloud-network-edge-device architecture of power IoT systems, the security monitoring of each node is characterized by isolated data and independent operations. Under this monitoring model, network traffic monitoring and terminal operation detection data fail to achieve deep interoperability and correlation, resulting in a single and fragmented dimension of security analysis. When a terminal is subjected to protocol instruction tampering or covert channel attacks, the master station often only captures surface-level traffic fluctuations, failing to obtain crucial contextual information such as abnormal processes and resource usage within the terminal. Furthermore, due to the massive number of power terminals, the master station's network traffic monitoring faces immense computational pressure, resulting in insufficient early warning timeliness. This lack of comprehensive data correlation and computation makes it difficult for existing systems to accurately identify and trace deep-seated security risks originating from power terminals in complex cross-regional and cross-domain attacks. For example, Chinese patent CN113569236A discloses an IoT terminal security monitoring and protection method aimed at solving the security risk problem in scenarios with massive heterogeneous terminal access. Its core idea is to construct a terminal security monitoring system in a cloud-edge collaborative environment by combining device identification, behavior profiling, and dynamic protection. This method uses the terminal's MAC address as a unique identifier, combining bypass traffic fingerprinting and active probe fingerprint scanning. It then uses the XGBoost algorithm to train a classification model for automatic terminal identification and type determination. However, this approach relies heavily on network traffic and fingerprint data for analysis, making it difficult to detect attacks exploiting vulnerabilities in business systems. Attackers can launch attacks even when network behavior appears normal, leading to missed alerts. All terminal data is uploaded to the cloud for unified analysis, neglecting the real-time computing capabilities of the edge, resulting in high cloud computing load and response latency. Furthermore, it cannot immediately block high-risk attacks, leaving a time window vulnerability in security response.
[0016] For example, Chinese patent CN112261041A discloses a multi-level distributed monitoring and anti-penetration system for power terminals, aiming to solve the problems of insufficient dimensions, poor universality, and inability to effectively block attacks in existing power terminal security monitoring. Its core idea is to construct a multi-level distributed security monitoring framework to simultaneously perform security perception at the terminal, network, and service levels, and implement differentiated anti-penetration strategies based on attack hazard assessment results. The system consists of two main parts: a security monitoring and anti-penetration linkage component and a multi-level distributed security monitoring and anti-penetration model framework for power terminals. This solution uses IP addresses as terminal identifiers. However, in power scenarios, service terminals often access the system through edge devices, and IP addresses may change dynamically or undergo NAT translation, making it impossible to establish a stable association between service terminals and edge access devices, and difficult to achieve data association between terminals, networks, and services. This solution adopts a centralized processing architecture, with all terminal perception data uploaded to a central module for analysis and modeling, without considering the computing power and real-time response requirements of the edge side. This deficiency leads to a high computational load and high response latency for the central node when facing a massive number of terminals, making it difficult to achieve immediate blocking of high-risk attacks.
[0017] Therefore, to address the insufficient cloud-edge collaborative analysis in power terminal security monitoring, it is necessary to research multi-level node monitoring data fusion and perception analysis technology for the power Internet of Things (IoT) to improve the monitoring and analysis capabilities for complex cross-regional and cross-domain high-stealth attacks on power terminals. This invention addresses the problems in existing technologies, such as isolated power terminal operation and business traffic data, and insufficient cloud-edge computing collaboration leading to low accuracy in cross-domain attack detection. It proposes a two-level cloud-edge collaborative perception architecture. At the edge, by collecting business terminal identifiers and operational violation data, an edge-end identifier association is established, and a terminal threat score is calculated in real time based on a dynamic baseline, enabling millisecond-level local handling of high-risk terminals. At the cloud, business traffic characteristics are analyzed through mirroring, and a full-element terminal perception vector is constructed based on the identifier association. The terminal is then classified and graded according to its score. Medium-risk terminals are analyzed using a temporal topology fusion network, while low-risk terminals are included in the cold data processing flow. This invention innovatively introduces a cross-modal temporal attention mechanism to explicitly model the attack characteristics where operational anomalies precede traffic anomalies, significantly improving the perception capability against high-penetration threats across regions and domains to power terminals.
[0018] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0019] Example 1 This invention provides a cloud-edge collaborative method for cross-domain threat perception of power terminals, as illustrated in the schematic diagram below. Figure 1 As shown, the method includes: Step S110: Receive the identifier mapping relationship, the operating status characteristics, and the edge threat score reported by the edge terminal; the identifier mapping relationship is used to indicate the association relationship between the edge terminal and each service terminal accessing the edge terminal; Step S120: If it is determined that the edge threat score of an edge terminal is within the medium risk threshold range, the monitored service traffic is analyzed to extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic. Step S130: Based on the identifier mapping relationship, associate the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identifier to construct a perception vector; Step S140: Input the perception vector into the temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal.
[0020] In this example implementation, the executing entity is the cloud master station (the master station side of the power system). It first receives the identifier mapping relationship, operational status characteristics, and edge threat score reported by the edge terminals. The identifier mapping relationship indicates the association between the identifiers of each service terminal accessing the edge terminal and the edge terminal's own edge terminal identifier. Specifically, the edge terminal extracts the unique DeviceID encoding of each service terminal in the IoT management platform by mirroring and parsing the northbound MQTT service traffic in its network interface card. It then binds these DeviceIDs with the edge terminal's own EdgeID to form...<EdgeID,DeviceID1,DeviceID2,…,DeviceIDn> The mapping relationship table is then reported to the cloud main station. At the same time, the edge terminals continuously collect their own operating status characteristics (such as CPU utilization, memory utilization, unauthorized external connections, etc.), and calculate the edge threat score of each business terminal using a dynamic baseline statistical method and report it together.
[0021] After obtaining the above data, the cloud-based master station determines that if any edge terminal's edge threat score falls within the medium-risk threshold range (e.g., between 60 and 99), it indicates that the edge terminal exhibits minor anomalies or occasional violations, requiring further assessment to determine if a highly covert attack with a long latency period is present. At this point, the cloud-based master station uses a traffic mirroring probe deployed on the business system side to perform bypass mirroring analysis on the monitored business traffic. This probe seamlessly acquires all business traffic data entering and leaving the master station via port mirroring, extracting network 5-tuple information, traffic statistics, and business characteristics such as MQTT (Message Queuing Telemetry Transport) protocol function code sequences and topic attribute values from each data packet using deep packet inspection (DIP) technology. Simultaneously, it parses the DeviceID of the business terminal to which the traffic belongs from the MQTT topic characteristics. Thus, the cloud-based master station obtains data streams from two different perception domains: one from the operational status characteristics and edge threat scores reported by the edge terminals, and the other from the business traffic characteristics mirrored and analyzed by the master station.
[0022] Based on this, the cloud master station, according to the received identifier mapping relationship, spatiotemporally aligns and associates traffic features containing the same DeviceID with the corresponding edge terminal operating status features to construct a full-element perception vector (perception vector) for the power terminal. Subsequently, the cloud master station inputs this full-element perception vector into a temporal topology fusion constraint network (TTF-Net) for in-depth threat analysis. This network takes the terminal's full-element perception vector as input and models the cross-modal temporal association risk and cross-node topology propagation risk of a single terminal node through a unified end-to-end architecture, ultimately outputting the threat identification result of the edge terminal. Through the above-mentioned edge-cloud bidirectional identifier association mechanism, this invention solves the inherent defects of existing technologies that cannot adapt to power scenarios such as NAT (Network Address Translation) and dynamic IP due to the use of IP addresses as identifiers. It realizes the spatiotemporal alignment and semantic association between business traffic features and terminal operating status features, forming a complete profile of terminal behavior. At the same time, by sinking the threat scoring calculation to the edge side and establishing a hierarchical processing mechanism, high-risk terminals can be handled locally at the edge side at the millisecond level, effectively eliminating the time window vulnerability of security response.
[0023] In some implementations, the step of inputting the perception vector into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal includes: Multi-head attention fusion is performed on the perception vectors to obtain fused features; The fused features are subjected to iterative convolutional processing layer by layer to obtain the propagation features; By applying nonlinear activation to the fusion feature and the propagation feature respectively, the temporal risk probability and the propagation risk probability are obtained; Based on the fusion features, the temporal risk probability, and the propagation risk probability, the threat identification result of the edge terminal is determined; The multi-head attention fusion is a feature fusion based on the temporal offset between the running state features and the business traffic features in the perception vector.
[0024] In this example implementation, such as Figure 2 As shown, the temporal topology fusion constraint network includes a multi-head attention unit, a graph convolution unit, a first classification unit, and a second classification unit. During threat analysis of the perceptual vector input into this network, the perceptual vector is first processed by the multi-head attention unit through multi-head attention fusion. Specifically, the operational state feature sequence and the business traffic feature sequence in the perceptual vector are projected onto a unified low-dimensional embedding space to construct the query matrix Q, key matrix K, and value matrix V required for the attention mechanism. Then, the temporal offset (e.g., sampling time difference) between the operational state features and business traffic features in the perceptual vector is encoded. Since the terminal operational state feature sequence and the business traffic feature sequence originate from different perceptual domains and the sampling time points may be offset, a learnable temporal offset encoding function is used to encode the relative time offset between the operational state features and traffic behavior features at different time steps, generating a temporal offset vector. The coupling strength between the query matrix and the key matrix at corresponding time steps is calculated based on the temporal offset vector. This coupling strength reflects both feature similarity and temporal causal lag. Subsequently, attention weights are calculated, and finally, the attention weights are applied to the corresponding business traffic features to obtain the fused features. This cross-modal temporal correlation computation method can explicitly capture the causal lag relationship between abnormal operating states and abnormal traffic, thereby ensuring that the network can correlate attack features at misaligned time points. By learning multiple terminal attack semantics in parallel through a multi-head attention mechanism, the system can accurately locate data segments with collaborative threat characteristics in massive business traffic, using the terminal's own operating state as an index. Subsequently, the fused features are iteratively convolved layer by layer through graph convolutional units to obtain propagation features. For cross-node topology propagation risks, the diffusion process of attack risks in the power grid topology is modeled iteratively through graph convolutional layers. For a single terminal node, after fusing and analyzing the temporal correlation information of operating state and business traffic, the probability of temporal risk is determined by nonlinear activation of the fused features through the first classification unit; the probability of propagation risk is obtained by nonlinear activation of the propagation features through the first classification unit, and then the comprehensive information of the fused features, the probability of temporal risk, and the probability of propagation risk is classified by the second classification unit to determine the threat identification result of the edge terminal. The above process enables the present invention to provide early warning of attack timing characteristics where operational anomalies precede traffic anomalies, upgrading from traditional passive response to proactive defense, effectively making up for the shortcomings of existing technologies that can only perform post-attack detection.
[0025] In some implementations, the multi-head attention fusion of the perceptual vectors to obtain fused features includes: The running status features in the perception vector are linearly transformed to generate a query matrix; the business traffic features in the perception vector are linearly transformed to obtain a key matrix and a value matrix; The temporal offset between the running state features and the business traffic features in the perception vector is encoded to obtain the temporal offset vector; The coupling strength between the query matrix and the corresponding time step in the key matrix is calculated based on the time offset vector, and then the attention weight is determined. The attention weights are applied to the corresponding business traffic features to obtain the fused features.
[0026] In this example implementation, such as Figure 2 As shown, the multi-head attention unit includes a linear transformation layer, a temporal coding layer, and an attention fusion layer. When the cloud master station performs feature fusion based on the temporal offset between the running state features and business traffic features in the perception vector, it first performs a linear transformation on the running state features in the perception vector to generate a query matrix; then it performs a linear transformation on the business traffic features in the perception vector to obtain a key matrix and a value matrix, which are mapped to the same low-dimensional embedding space. The query matrix originates from the running state feature sequence, and the key and value matrices originate from the business traffic feature sequence. This asymmetric construction method ensures the attention logic of retrieving relevant abnormal segments from the business traffic using running state anomalies as an index. Subsequently, the temporal coding layer encodes the temporal offset between the running state features and business traffic features in the perception vector to obtain a temporal offset vector. The temporal offset vector encodes the relative time difference between the p-th running state time step and the q-th traffic time step using a learnable temporal offset coding function. p and q belong to the same risk perception / monitoring time window, allowing the network to freely associate attack features at misaligned time points, such as a running anomaly at the 5th minute and lateral movement traffic appearing only at the 10th minute. Based on this temporal offset vector, the network calculates the coupling strength between the query matrix and the key matrix at corresponding time steps. This coupling strength is then normalized using Softmax to obtain the attention weight. A larger weight indicates that the corresponding business traffic feature is more likely to have a causal relationship with previous operational anomalies. Finally, the attention fusion layer applies the attention weights to the corresponding value matrix (i.e., business traffic features), performing a weighted summation of the business traffic features at each time step to obtain the fused feature. This fused feature, using operational anomalies as clues, accurately locates data segments with collaborative threat characteristics within massive amounts of business traffic.
[0027] In some implementations, the service traffic characteristics include: Network quintuple information, service traffic statistical characteristics, and service attribute characteristics; The operational status characteristics include at least one of the following: CPU utilization, memory utilization, unauthorized external connections, unauthorized process execution, high-risk port opening, unauthorized peripheral device access, and critical configuration changes.
[0028] In this example implementation, the edge terminal continuously collects its own operational status characteristics to construct its own operational status feature vector. These operational status characteristics may include at least one of seven categories of operational indicators: CPU utilization, memory utilization, unauthorized external connections, unauthorized process execution, high-risk port opening, unauthorized peripheral device access, and critical configuration changes. All collected data is timestamped to form an operational status sequence. The power terminal service traffic is mirrored and analyzed on the main station side, and a terminal service traffic feature vector is constructed in real time. , For sampling timestamps, For the nth service traffic feature, the service traffic feature may include: a five-tuple: source IP, destination IP, source port, destination port, and transport layer protocol type; Traffic statistics features include packet length, packet interval, bytes per second, and connection duration. Service attribute features include MQTT protocol function code sequences and service attribute values. Among these, the DeviceID of the service terminal can be extracted through MQTT topic feature parsing. Based on the reported edge identifier mapping table, the power terminal operating status feature vector and service traffic feature vector can be associated and concatenated to construct a comprehensive perception vector for the service terminal.
[0029] Upon receiving the edge threat score, the cloud-based main station performs tiered classification processing on the edge terminal's edge threat score. If the edge threat score falls into the high-risk range (e.g., S≤59), immediate local security measures are triggered on the edge terminal. If the edge threat score of an edge terminal is determined to be in the low-risk range (e.g., S=100), only the edge terminal's operational status characteristics and edge threat score are transferred to the cold data storage area, without further real-time analysis. If the edge threat score of an edge terminal is determined to be in the medium-risk range (60≤S≤99), a continuous and in-depth risk assessment of the terminal is conducted in conjunction with the full-element perception vector.
[0030] For medium-risk terminals, the main station employs a Temporal-Topology FusionConstraint Network (TTF-Net) for refined detection. The network structure is as follows: Figure 2 As shown, using the terminal's full-element perception vector and the power grid topology as input, and through a unified end-to-end architecture, we model the cross-modal temporal correlation risk and cross-node topology propagation risk of single-service terminal nodes, and deeply explore cross-regional and cross-domain penetration attacks with long incubation periods and high concealment.
[0031] A cross-modal mapping based on a multi-head attention mechanism is employed, using the operational state encoding vector as Query(Q) and the traffic feature encoding vectors as Key(K) and Value(V) to explicitly model the temporal offset relationship between operational anomalies and traffic anomalies. This is because the operational state feature sequence of the edge terminal... With business traffic feature sequence Originating from different perceptual domains, the two are first projected into a unified low-dimensional embedding space using a linear transformation layer to construct the query, key, and value matrix required by the attention mechanism.
[0032] The linear transformation method is as follows:
[0033] in, , , These are the query matrix, key matrix, and value matrix, respectively. , , These are the weight matrices for the attention heads corresponding to the query, key, and value, respectively. , These are the running status feature vector and the service traffic feature vector in the perception vector.
[0034] To capture the causal lag between abnormal power terminal operating status and abnormal service traffic on the time axis, such as the occurrence of cross-regional lateral movement traffic several minutes after a terminal escalation event, this invention introduces a relative temporal offset operator based on dot product attention.
[0035] The associated calculation method is: the computing terminal's first p The running state characteristics of the first time step and the second time step q The coupling strength between traffic behavior features at each time step:
[0036]
[0037] in , For the first p The running state characteristics of the first time step and the second time step q The coupling strength between the business traffic characteristics at each time step For the first p The running state characteristics of the first time step and the second time step The coupling strength between the business traffic characteristics of each time step; Represents the coding function of a learnable temporal coding layer; For the first p The running state characteristics of the first time step and the second time stepq The temporal offset vector between the business traffic characteristics of each time step; The dimension of the key is T, and the superscript T denotes the transpose of the matrix. For matrix The Middle q Each time step corresponds to a vector. For matrix The Middle p Each time step corresponds to a vector.
[0038] A multi-head attention mechanism is employed to learn various terminal attack semantics in parallel, and attention weights are applied to traffic features. The attention weighting formula is as follows:
[0039]
[0040] In the formula, As a feature of fusion, For the first p The running state characteristics of the first time step and the second time step q Attention weights among business traffic features at each time step For matrix The Middle q Each time step corresponds to a vector. L This represents the current risk perception time window length.
[0041] In this way, the system can use the operating status of the edge terminal as an index to accurately locate data segments with collaborative threat characteristics in massive business traffic.
[0042] For a single terminal node, the timing correlation information between its operating status and service traffic is analyzed in a fusion manner, and the terminal timing risk is calculated as follows:
[0043] in, The probability of timing risk for business terminals. 、 for Trainable network parameters; 、 for Trainable network parameters, It is an exponential linear activation function. This is the activation function used to compress the output to the [0,1] range. The larger the value, the higher the risk of abnormal behavior of the terminal node.
[0044] In some implementations, the step of performing layer-by-layer iterative convolution processing on the fused features to obtain propagation features includes: According to the business communication path of the edge terminal in the terminal association graph, the fusion feature is subjected to layer-by-layer iterative convolution processing to obtain the propagation feature; The terminal association graph is constructed based on the business communication connection relationship of each power terminal in the power Internet of Things; the propagation weight between adjacent convolutional layers in the layer-by-layer iterative convolutional processing is determined based on the multi-dimensional registration attributes between adjacent nodes of the corresponding propagation segment in the terminal association graph.
[0045] In this example implementation, to further identify cross-terminal node coordinated attacks and cross-regional lateral attacks, relying solely on single-terminal time-series analysis may still miss low-frequency, slow-propagation behaviors. Therefore, this invention, considering the characteristics of the power grid cloud-edge-terminal hierarchical topology, further introduces a power grid topology propagation chain constraint mechanism to analyze cross-node topology propagation risks. Figure 2 As shown, the fused features are further processed using graph convolutional units. The power grid master station constructs a terminal association graph based on the power terminal access relationships and the business communication connection relationships of each terminal within the power Internet of Things. ,in, This represents a set of power terminal nodes, which include edge terminals and service terminals. This represents the set of edges between nodes, i.e., the business connection relationships between power terminals. Node attributes can include multi-dimensional registration attributes from the terminal's business application registration. The power grid master station can construct a corresponding terminal association graph based on currently monitored medium-risk terminals, or it can extract the association subgraph corresponding to the current medium-risk terminals from a pre-constructed terminal association graph corresponding to the entire network topology. For the power terminal / edge terminal nodes to be analyzed... i The graph convolutional unit uses the edge indices of the terminal association graph of the edge terminal to perform convolution processing on the fused features, completing one message aggregation and feature update. Specifically, the initial node feature vector... Defined as:
[0046] in, This represents the fusion feature of power terminal / edge terminal node i. For the power terminal / edge terminal to be analyzed, the graph convolutional unit performs iterative convolution processing on the fusion feature according to the edge index of the power terminal / edge terminal in the terminal association graph. In each layer of graph convolution, the feature vector of the current node is updated by aggregating the features of all its neighboring nodes. The propagation weights during aggregation (propagation weights of adjacent layers in the iterative convolution processing) are determined based on the multi-dimensional registration attributes between corresponding two nodes in the terminal association graph. The multi-dimensional registration attributes are the attribute features of the terminal when registering for business applications. For example, the multi-dimensional registration attributes include at least one of business affiliation, access gateway relationship, and transformer area affiliation. For example, the diffusion process of attack risk in the power grid topology is modeled iteratively by the following graph convolutional layer:
[0047] in, , The first , Propagation characteristics of each propagation layer Nodes in the terminal association graph The set of neighboring nodes, The risk propagation weights between nodes i and j in the terminal association graph. Let be the trainable transformation matrix of the graph convolutional unit. is the nonlinear activation function for the graph convolution unit.
[0048] Based on the power grid topology association characteristics of terminal nodes, propagation weights are used. Defined as:
[0049] in, , , These are the learnable weight parameters of the graph convolutional unit. This indicates whether nodes i and j in the terminal association diagram belong to the same business application. This indicates whether the network access of nodes i and j in the terminal association diagram belongs to the same secure access gateway. This indicates whether nodes i and j in the terminal association diagram are in the same power grid area. , , The values are all 0 or 1. For example, when nodes i and j belong to the same business application, ,otherwise, , and The values are similar.
[0050] The aforementioned propagation mechanism allows for the construction of propagation paths for attack risks within the terminal network. When multiple terminals have timing risk values... When a continuous growth trend is observed, an attack propagation chain can be formed. :
[0051] in, This represents the m-th attack propagation node / power terminal.
[0052] Mapping the propagated node features to scalar risk values yields the topology propagation risk score for the terminal nodes:
[0053] in, This represents the total number of layers in the attack propagation. The activation function for the propagation process can be trained with parameters. The probability of propagation risk at power terminals / edge terminals. For nodes i The propagation characteristics output in the last propagation layer.
[0054] In some implementations, determining the threat identification result of the edge terminal based on the fusion features, the temporal risk probability, and the propagation risk probability includes: The temporal risk and the propagation risk are weighted and fused to obtain the comprehensive risk value of the edge terminal; The fusion feature, the temporal risk probability, and the propagation risk probability are concatenated into vectors to form an enhanced feature; The enhanced features are nonlinearly mapped to obtain the risk category of the edge terminal; The risk category and overall risk value of the edge terminal are used as the threat identification result.
[0055] In this example implementation, such as Figure 2 As shown, the temporal topology fusion constraint network also includes a fusion unit. The fusion unit performs weighted fusion of the temporal risk probability and the propagation risk probability, that is, it fuses the temporal risk of the terminal node with the cross-node topology propagation risk. The final risk assessment calculation of the terminal is defined as follows:
[0056] in, Represented as the fusion coefficient, This represents the overall risk value for edge terminals.
[0057] Based on this, the cloud-based main station concatenates the fusion characteristics, temporal risk probability, and propagation risk probability of the edge terminal into an enhanced feature vector. :
[0058] Subsequently, the enhanced features are input into the second classification unit (deep classification network) for non-linear mapping, outputting the final terminal risk category. That is, the original fused features are concatenated with the two risk scores and input into the deep classification network for non-linear mapping. Features are then refined through gated linear units and subsequently input into the Softmax classification layer.
[0059]
[0060] in, Let be the risk probability distribution vector. This represents the probability that the edge terminal belongs to each risk category. , For the network parameters of the gated linear unit, , These are the network parameters for the Softmax classification layer. The final category is identified. Determined by the category corresponding to the highest probability: .
[0061] This example categorizes risks into at least one of the following: malicious scanning, brute-force attacks, replay attacks, sensitive data breaches, and normal operation. This example demonstrates the accurate identification of various terminal anomaly risks, including malicious scanning, brute-force attacks, replay attacks, and sensitive data breaches.
[0062] Existing technologies use IP addresses as unique identifiers for terminals. However, in power scenarios, service terminals often access the network through edge devices, and IP addresses may change dynamically or undergo NAT translation, making it impossible to establish a stable association between the service terminal and the edge access device. This invention establishes an edge-cloud bidirectional identifier association mechanism, mapping and associating service terminal identifiers with edge terminal identifiers. This enables end-to-end traceability from service flow to physical devices, solving the problem of cross-regional and cross-domain data association in power service scenarios.
[0063] Existing technologies employ a centralized processing architecture, where all terminal-perceived data is uploaded to a central module for analysis and modeling, failing to utilize edge computing power for real-time response. When dealing with massive numbers of terminals, the central node experiences heavy computational load and high response latency, making it unable to immediately block high-risk attacks. This invention addresses this by calculating terminal threat scores in real-time at the edge, triggering security measures directly at the edge for high-risk terminals, compressing response time to milliseconds, and eliminating the time window vulnerability for security responses.
[0064] Example 2 Based on the same inventive concept, this invention provides a cloud-edge collaborative cross-domain threat perception method for power terminals, applied to edge terminals, including: The monitored service traffic is analyzed to extract the service terminal identifiers of each service terminal accessing the edge terminal, and the service terminal identifiers are associated with the edge terminal to generate an identifier mapping relationship; The edge terminal collects its own operational status characteristics and calculates its edge threat score based on the degree of deviation of the operational status characteristics from the baseline. The identification mapping relationship, the operational status characteristics, and the edge threat score are reported to the cloud master station, so that the cloud master station can analyze the monitored service traffic, extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic, and associate the service traffic characteristics of each service terminal with the operational status characteristics of the edge terminal associated with the service terminal identifier based on the identification mapping relationship to form a perception vector, and input the perception vector into the temporal topology fusion constraint network for threat analysis.
[0065] In this example implementation, the execution entity is an edge terminal. The edge terminal first parses the service traffic of each service terminal accessing it, extracts the service terminal identifier of each terminal, and associates these identifiers with its own edge terminal identifier to generate an identifier mapping relationship. Specifically, the edge terminal obtains northbound service traffic from its network interface card (NIC) via port mirroring, performs deep parsing of MQTT protocol messages, extracts the unique DeviceID of each service terminal in the IoT management platform from the message subject characteristics, and simultaneously obtains its own unique EdgeID, binding the two together to form an identifier mapping relationship.<EdgeID,DeviceID1,DeviceID2,…,DeviceIDn> A formal identifier mapping table is used. This identifier mapping relationship solves the technical defects of existing technologies that rely solely on IP addresses as terminal identifiers, leading to identifier failure due to NAT translation or dynamic DHCP allocation, and the inability to establish a stable association between service terminals and edge access devices. Simultaneously, the edge terminal continuously collects its own operational status characteristics. These operational status characteristics include multiple operational indicators. The edge terminal maintains a dynamic baseline for each operational indicator, and the baseline mean and standard deviation are updated in real time through sliding window statistics of historical operational data. The edge terminal uses a dynamic baseline statistical method to calculate the terminal edge threat score (0-100 points) in real time; a higher score indicates a lower risk. The identifier mapping relationship, the operational status characteristics, and the edge threat score are reported to the cloud master station for threat analysis. For example, based on the edge threat score, edge terminals are divided into three levels: Low-risk terminals: S=100, indicating that all indicators are within the baseline range and there are no violations. Medium-risk terminals: 60≤S≤99, indicating minor anomalies or occasional violations, requiring further analysis. High-risk terminals: 1≤S≤ 59, indicating serious anomalies or continuous violations, possibly indicating intrusion, requiring immediate local handling.
[0066] In some implementations, after calculating the edge threat score, the method further includes: If an edge terminal is determined to have an edge threat score that falls within the high-risk range, local security measures will be taken against that edge terminal.
[0067] In this example implementation, if the edge threat score of an edge terminal falls within the high-risk range, for example, a score between 1 and 59, it indicates that the terminal exhibits serious anomalies or persistent violations, possessing a high probability of being compromised. The edge terminal immediately executes local security measures. Local security measures include, but are not limited to, immediate response operations such as process termination and / or high-risk port closure. For example, when the security monitoring probe of the edge terminal detects a violating process running, it directly issues a process termination command to the operating system to end the execution of the malicious process; when a high-risk port is detected open, it directly issues a port disabling command to block potential backdoor communication channels. The execution of the above-mentioned measures is completed entirely at the edge, without waiting for cloud command feedback. The response time can be compressed to less than 50 milliseconds, which is two orders of magnitude faster than the traditional cloud-based centralized measures mode (average response time of 5 to 10 seconds). This effectively eliminates the time window vulnerability of security response and prevents attackers from using decision-making delays to complete further penetration operations.
[0068] In some implementations, the process of constructing the identifier mapping relationship includes: The network interface card (NIC) service traffic of the edge terminal is obtained by port mirroring, and the service terminal identifier in the NIC service traffic is extracted by parsing the MQTT northbound service traffic protocol. The business terminal identifier is bound to the edge terminal identifier of the edge terminal itself to form the identifier mapping relationship.
[0069] In this example implementation, the security monitoring probe of the edge terminal extracts northbound MQTT protocol messages from the mirrored traffic. MQTT is a standard lightweight messaging protocol for business data interaction between the power IoT edge IoT agent and the IoT management platform. The topic field of the message carries the DeviceID information of the service terminal from which the message originates. The probe extracts the corresponding service terminal identifier DeviceID from each message by parsing the topic characteristics of the MQTT message. Subsequently, the probe binds this DeviceID to the edge terminal's own EdgeID, forming an identifier mapping relationship. Since multiple service terminals are typically connected to an edge terminal, the final identifier mapping relationship is represented as a one-to-many mapping table structure. This mapping table uses EdgeID as the primary key and a list of DeviceIDs as attribute values, fully recording the subordinate association relationships between the edge terminal and its subordinate service terminals.
[0070] In some implementations, calculating the edge threat score of the edge terminal based on the degree of deviation of the operating state characteristics from the baseline includes: For each operational indicator in the operational status characteristics, calculate the deviation of the current measured value from the corresponding baseline, and determine the abnormal score of the operational indicator based on the deviation. The abnormal scores of each operational indicator within the operational status feature are weighted and summed to obtain a comprehensive abnormal value; The comprehensive outlier values are numerically normalized to obtain the edge threat score of the edge terminal; The baseline for each operational indicator is determined based on the statistical measures of its historical values.
[0071] In this example implementation, operational status characteristics may include multiple operational metrics such as CPU utilization, memory utilization, unauthorized external connection events, unauthorized process execution events, high-risk port opening events, unauthorized peripheral device access events, and critical configuration change events. For each operational metric, an anomaly score is calculated based on the degree of deviation between the current measured value and the baseline. Taking CPU utilization as an example: Let the baseline mean of CPU utilization be... The standard deviation is The current measured value is Define the deviation. :
[0072] in, It is a very small constant.
[0073] Abnormal terminal indicator scores for:
[0074] in, It is the first Weighting factors for each operating indicator.
[0075] For discrete indicators, such as unauthorized external connections or unauthorized processes, a high anomaly score is directly assigned if an unauthorized behavior occurs; otherwise, the score is 0.
[0076] The comprehensive threat score is processed using a weighted summation and normalization. Let the outlier scores for the seven categories of indicators be... The corresponding weights are respectively .
[0077] Comprehensive outliers for:
[0078] The edge threat score is:
[0079] When E≥100, S=0; when E=0, S=100.
[0080] Edge threat scoring of edge terminals and running state feature vector Report to the main station.
[0081] For example, such as Figure 3 As shown, this invention provides a cloud-edge collaborative power terminal threat perception method. At the edge terminal, business terminal identifiers and operational violations are collected. The extracted identifiers are associated with the edge terminal identifiers. The edge terminal calculates its own threat score and reports the edge terminal identifier association and threat score to the cloud master station. The master station performs mirror parsing to extract business traffic features. Based on the edge terminal identifier association, the traffic features carrying the business terminal identifiers are associated with the edge terminal score to form a full-element perception vector for the power terminal. For edge terminals with high threat scores, edge-side security measures are directly implemented. For terminals with medium threat scores, time-series analysis is performed on the full-element perception vector to identify abnormal attack behaviors. For terminals with low-risk edge terminal scores, they are included in the cold data processing flow and not subjected to in-depth analysis. This invention achieves full-element perception data association analysis for power terminals through edge device identifier association and computational collaboration, improving the monitoring accuracy and early warning efficiency for power terminal attack threats.
[0082] This invention establishes an edge-cloud bidirectional identifier association mechanism to map and associate service terminal identifiers with edge terminal identifiers, thus solving the inherent defects of existing power scenarios, such as NAT translation and dynamic IP, which are unable to adapt to these limitations due to the use of IP addresses as identifiers. Based on this identifier association, this invention achieves spatiotemporal alignment and semantic association between service traffic characteristics and terminal operating status characteristics, constructing a full-element perception vector for power terminals and forming a complete profile of terminal behavior.
[0083] Experimental example: A semi-simulation experiment was conducted based on a distributed photovoltaic experimental environment. The terminals included edge terminal devices such as transformer substation fusion terminals and concentrators, as well as service terminals such as photovoltaic inverters, smart circuit breakers, low-voltage branch monitoring terminals, temperature and humidity sensors, smoke sensors, and water immersion sensors. Communication between the edge terminals and the main station was achieved using MQTT communication based on Mosquitto software.
[0084] The penetration test simulated an attack on the main station's distributed photovoltaic system through an edge device. Risk identification was performed using both the present invention and the traditional traffic-based anomaly risk detection method. The results are shown in Tables 1 and 2.
[0085] Table 1 Performance of the Anomaly Risk Identification Method of the Invention
[0086] Table 2. Performance comparison between the method of the present invention and the method based on traffic anomaly risk identification.
[0087] As shown in Tables 1-2, the present invention achieves an accuracy rate of 96.8% in identifying four types of terminal anomaly risks: malicious scanning, brute-force attacks, replay attacks, and sensitive data leakage, with a false negative rate of 3.2% and a false positive rate of 1.7%. Compared with anomaly risk identification methods that rely on single business traffic, the accuracy rate is improved by 4.5 percentage points, demonstrating excellent performance in simulated cross-regional penetration attack scenarios.
[0088] This invention breaks through the limitations of existing centralized processing architectures by innovatively pushing threat scoring calculation down to the edge and establishing a three-tiered collaborative processing mechanism: real-time edge response, in-depth cloud analysis, and tiered cold data storage. At the edge, threat scores for seven operational metrics are calculated in real-time based on a dynamic baseline statistical algorithm, directly triggering local blocking and isolation for high-risk terminals. Response time is reduced to less than 50ms, an improvement of two orders of magnitude compared to the centralized cloud processing model (average response time 5-10 seconds), effectively eliminating the time window vulnerability for security response.
[0089] Meanwhile, for a massive number of power terminals, low-risk terminals (accounting for approximately 98%) are downgraded to cold data, with only metadata stored. Based on a scale of millions of terminals, the number of terminals requiring real-time analysis daily has decreased from 1 million to 20,000, reducing the computing power requirement for deep cloud analysis by 80%, storage overhead by 60%, and significantly improving the overall analytical performance of the system.
[0090] This invention proposes a time-series topology-constrained network that innovatively introduces cross-modal time-series correlation computation based on multi-head attention to address the attack timing characteristics where operational anomalies precede traffic anomalies. The model can explicitly capture the causal lag relationship between operational anomalies and traffic anomalies. This technology effectively overcomes the limitations of existing methods that only perform post-attack detection and cannot provide early warning of highly covert attacks, upgrading from passive response to proactive defense and significantly improving the power system's ability to prevent highly covert cross-regional and cross-domain attacks.
[0091] Example 3 Based on the same inventive concept, the present invention also provides a cloud-based main station, comprising: The main station traffic mirroring probe is used to receive the identifier mapping relationship, the operating status characteristics, and the edge threat score reported by the edge terminal. The identifier mapping relationship is used to indicate the association relationship between the edge terminal and each service terminal accessing the edge terminal. It is also used to analyze the monitored service traffic and extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic if it is determined that the edge threat score of an edge terminal is within the medium risk threshold range. Based on the identifier mapping relationship, the service traffic characteristics of each service terminal are associated with the operating status characteristics of the edge terminal associated with the service terminal identifier to construct a perception vector. The terminal anomaly risk deep analysis engine is used to input the perception vector into a temporal topology fusion constraint network for threat analysis, and obtain the threat identification result of the edge terminal.
[0092] In one possible implementation, the temporal topology fusion constraint network includes a multi-head attention unit, a graph convolution unit, a first classification unit, and a second classification unit; The multi-head attention unit is used to perform multi-head attention fusion on the perception vector to obtain fused features; The graph convolutional unit is used to perform layer-by-layer iterative convolution processing on the fused features to obtain the propagation features; The first classification unit is used to perform nonlinear activation on the fusion feature and the propagation feature respectively to obtain the temporal risk probability and the propagation risk probability; The second classification unit is used to determine the threat identification result of the edge terminal based on the fusion features, the temporal risk probability, and the propagation risk probability; The multi-head attention fusion is a feature fusion based on the temporal offset between the running state features and the business traffic features in the perception vector.
[0093] In one possible implementation, the multi-head attention unit includes a linear transformation layer, a temporal coding layer, and an attention fusion layer; The linear transformation layer is used to: perform a linear transformation on the running state features in the perception vector to generate a query matrix; and also to perform a linear transformation on the business traffic features in the perception vector to obtain a key matrix and a value matrix. The temporal coding layer is used to: encode the temporal offset between the running state features and the business traffic features in the perception vector to obtain a temporal offset vector; The attention fusion layer is used to: calculate the coupling strength between the query matrix and the corresponding time step in the key matrix based on the temporal offset vector, and then determine the attention weights; The attention fusion layer is also used to: apply the attention weights to the corresponding business traffic features to obtain fused features.
[0094] In one possible implementation, the formula for calculating the fusion feature is:
[0095]
[0096]
[0097]
[0098] in, As a feature of fusion, For the first p The running state characteristics of the first time step and the second time step q Attention weights among business traffic features at each time step For the first p The running state characteristics of the first time step and the second time step q The coupling strength between the business traffic characteristics at each time step For the first p The running state characteristics of the first time step and the second time step The coupling strength between the business traffic characteristics of each time step; , , These are the query matrix, key matrix, and value matrix, respectively. , , These are the weight matrices for the attention heads corresponding to the query, key, and value, respectively. , These are the running state feature vector and the service traffic feature vector, respectively, in the perception vector. , Each is a matrix , The Middle q Each time step corresponds to a vector. For matrix The Middle p Each time step corresponds to a vector. L This represents the current risk perception time window length. For the first p The running state characteristics of the first time step and the second time step q The temporal offset vector between the business traffic characteristics of each time step; The dimension of the key is T, and the superscript T denotes the transpose of the matrix.
[0099] In one possible implementation, the graph convolution unit is specifically used for: According to the business communication path of the edge terminal in the terminal association graph, the fusion feature is subjected to layer-by-layer iterative convolution processing to obtain the propagation feature; The terminal association graph is constructed based on the business communication connection relationship of each power terminal in the power Internet of Things; the propagation weight between adjacent convolutional layers in the layer-by-layer iterative convolutional processing is determined based on the multi-dimensional registration attributes between adjacent nodes of the corresponding propagation segment in the terminal association graph.
[0100] In one possible implementation, the multi-dimensional registration attributes include at least one of service affiliation, access gateway affiliation, and station affiliation.
[0101] In one possible implementation, the formula for calculating the propagation characteristic is:
[0102]
[0103] in, , The first , The propagation characteristics of each convolutional layer Nodes in the terminal association graph The set of neighboring nodes, The risk propagation weights between nodes i and j in the terminal association graph. Let be the trainable transformation matrix of the graph convolutional unit. The nonlinear activation function of the graph convolutional unit, , , These are the learnable weight parameters of the graph convolutional unit. This indicates whether nodes i and j in the terminal association diagram belong to the same business application. This indicates whether the network access of nodes i and j in the terminal association diagram belongs to the same secure access gateway. This indicates whether nodes i and j in the terminal association diagram are in the same power grid area. , , The values are all 0 or 1.
[0104] In one possible implementation, the temporal topology fusion constraint network further includes a fusion unit; The fusion unit is used to: perform weighted fusion of the temporal risk and the propagation risk to obtain the comprehensive risk value of the edge terminal; The second classification unit is specifically used to: concatenate the fusion feature, the temporal risk probability, and the propagation risk probability into vectors to form an enhanced feature; input the enhanced feature into the classification unit for nonlinear mapping and output the risk category of the edge terminal; and use the risk category and comprehensive risk value of the edge terminal as the threat identification result.
[0105] In one possible implementation, the terminal anomaly risk deep analysis engine is also used for: If the edge threat score is determined to be in the low-risk range, the operating status characteristics and edge threat score of the edge terminal are transferred to the cold data storage area.
[0106] In one possible implementation, the service traffic characteristics include: Network quintuple information, service traffic statistical characteristics, and service attribute characteristics; The operational status characteristics include at least one of the following: CPU utilization, memory utilization, unauthorized external connections, unauthorized process execution, high-risk port opening, unauthorized peripheral device access, and critical configuration changes.
[0107] Example 4 Based on the same inventive concept, the present invention also provides an edge terminal, including a terminal security monitoring probe, wherein the terminal security monitoring probe includes: The mapping generation unit is used to parse the monitored service traffic, extract the service terminal identifiers of each service terminal accessing the edge terminal, and associate the service terminal identifiers with the edge terminal to generate an identifier mapping relationship. The threat calculation unit is used to collect its own operational status characteristics and calculate the edge threat score of the edge terminal based on the degree of deviation of the operational status characteristics from the baseline. The reporting unit is used to report the identifier mapping relationship, the operating status characteristics, and the edge threat score to the cloud master station, so that the cloud master station can analyze the monitored service traffic, extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic, and associate the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identifier based on the identifier mapping relationship to form a perception vector, and input the perception vector into the temporal topology fusion constraint network for threat analysis.
[0108] In one possible implementation, the threat calculation unit is specifically used for: For each operational indicator in the operational status characteristics, calculate the deviation of the current measured value from the corresponding baseline, and determine the abnormal score of the operational indicator based on the deviation. The abnormal scores of each operational indicator within the operational status feature are weighted and summed to obtain a comprehensive abnormal value; The comprehensive outlier values are numerically normalized to obtain the edge threat score of the edge terminal; The baseline for each operational indicator is determined based on the statistical measures of its historical values.
[0109] In one possible implementation, the terminal security monitoring probe further includes: The security handling unit is used to perform local security handling on the corresponding edge terminal if the edge threat score is determined to be in the high-risk range.
[0110] In one possible implementation, the mapping generation unit is specifically used for: The network interface card (NIC) service traffic of the edge terminal is obtained by port mirroring, and the service terminal identifier in the NIC service traffic is extracted by parsing the MQTT northbound service traffic protocol. The business terminal identifier is bound to the edge terminal identifier of the edge terminal itself to form the identifier mapping relationship.
[0111] Example 5 Based on the same inventive concept, this invention also provides a cloud-edge collaborative method for cross-domain threat perception of power terminals, including: The edge terminal parses the monitored service traffic, extracts the service terminal identifiers of each service terminal accessing the edge terminal, and associates the service terminal identifiers with the edge terminal to generate an identifier mapping relationship; The edge terminal collects its own operating status characteristics and calculates the edge threat score based on the degree of deviation of the operating status characteristics from the baseline; the identification mapping relationship, the operating status characteristics, and the edge threat score are reported to the cloud master station. When the cloud master station determines that the edge threat score of an edge terminal is within the medium risk threshold range, it analyzes the monitored business traffic and extracts the business traffic characteristics of each business terminal and the business terminal identifier carried in the business traffic. Based on the identifier mapping relationship, the cloud master station associates the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identifier to construct a perception vector; and inputs the perception vector into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal.
[0112] Example 6 Based on the same inventive concept, the present invention also provides a cloud-edge collaborative power terminal cross-domain threat perception system, including: the cloud master station of embodiment 3 and / or the edge terminal of embodiment 4.
[0113] For example, such as Figure 4 As shown, the system includes a cloud-based main station and edge terminals and service terminals (such as service terminal 1, service terminal 2, ..., service terminal N) on the edge side. The service applications on the edge terminals perform load balancing and business communication with the power service system through a secure access gateway. The edge terminals include terminal security monitoring probes. The cloud-based main station includes a main station traffic mirroring probe (which can be placed at the load balancer), a terminal anomaly risk in-depth analysis engine, and a monitoring visualization module. Details are as follows: 1) Terminal security monitoring probe Supports the association of edge terminal and service terminal identifiers. By continuously mirroring and collecting the service terminal identifier DeviceID from the edge terminal, and associating this identifier with the edge terminal EdgeID generated by the terminal security monitoring probe, an identifier association mapping table is formed.
[0114] It supports continuous collection of data on seven types of operational violations, including CPU utilization, memory utilization, unauthorized external connections, unauthorized processes, high-risk ports, unauthorized peripherals, and critical configuration changes. It also supports real-time calculation of endpoint threat scores on the edge terminal using a dynamic baseline statistical algorithm, and performs tiered processing based on the score results.
[0115] It supports the immediate triggering of end-side handling strategies for high-risk terminals, such as process killing and port closure.
[0116] It supports reporting the identifier association mapping table and terminal running status data to the main station system.
[0117] Deployed at power edge terminals, such as converged terminals and concentrators, it is responsible for real-time data collection and preliminary analysis on the terminal side.
[0118] 2) Main site traffic mirroring probe It supports association of main station side terminal identifiers, real-time mirror parsing of power terminal service traffic, and extraction of service terminal identifier DeviceID.
[0119] It supports the parsing of multi-dimensional traffic characteristics of terminals, including five-tuple information, network statistics, and service characteristics. The parsing results are output in time series format and matched with the identifier mapping relationship reported by the terminal security monitoring probe, realizing the spatiotemporal alignment of service traffic data and terminal operating status data, and providing support for the construction of a full-element perception vector.
[0120] The bypass is deployed on the power station's main business system side, and obtains all terminal business traffic data entering and leaving the main station seamlessly through port mirroring.
[0121] 3) Terminal Anomaly Risk Deep Analysis Engine It supports classifying and grading terminal risks, and for high-risk terminals, it can immediately trigger cloud-based blocking and isolation, such as IP blocking and revocation of access certificates; For medium-risk terminals, it supports refined detection based on the TTF-Net algorithm. It supports accurate identification of five states: malicious scanning, brute-force attacks, replay attacks, sensitive data leakage, and normal operation.
[0122] For low-risk terminals, the terminal's perceived data is stored in the cold data area and is not subjected to in-depth analysis.
[0123] Deployed on the cloud-based main site, it serves as the core of the system's intelligent analysis.
[0124] 4) Monitoring Visualization Module It supports statistics on power terminal assets and risk alarms, such as the total number of alarm risks, the distribution of risk types, and the time trend of risks. It also supports the display of terminal risk lists.
[0125] It supports cloud-based security measures for high-risk endpoints, such as one-click execution of blocking / isolation commands to handle endpoint threats.
[0126] Deployed in a cloud-based security operations and maintenance center, it provides a human-computer interaction interface and global situational awareness capabilities.
[0127] Example 7 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). An electronic device readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both the built-in storage medium within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the cloud-edge collaborative power terminal cross-domain threat perception method described in the above embodiments.
[0128] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0129] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0130] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0131] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims pending approval.
Claims
1. A cloud-edge collaborative power terminal cross-domain threat perception method, characterized in that, Applied to cloud-based main sites, including: The system receives the identifier mapping relationship, the operating status characteristics, and the edge threat score reported by the edge terminal; the identifier mapping relationship is used to indicate the association relationship between the edge terminal and each service terminal accessing the edge terminal. If it is determined that the edge threat score of an edge terminal is within the medium risk threshold range, the monitored service traffic is analyzed to extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic. Based on the identification mapping relationship, the service traffic characteristics of each service terminal are associated with the operating status characteristics of the edge terminal associated with the service terminal identification to construct a perception vector; The perception vector is input into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal; The step of inputting the perception vector into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal includes: Multi-head attention fusion is performed on the perception vectors to obtain fused features; The fused features are subjected to iterative convolutional processing layer by layer to obtain the propagation features; By applying nonlinear activation to the fusion feature and the propagation feature respectively, the temporal risk probability and the propagation risk probability are obtained; Based on the fusion features, the temporal risk probability, and the propagation risk probability, the threat identification result of the edge terminal is determined; The multi-head attention fusion is a feature fusion based on the temporal offset between the running state features and the business traffic features in the perception vector; The process of performing multi-head attention fusion on the perception vector to obtain fused features includes: The running status features in the perception vector are linearly transformed to generate a query matrix; the business traffic features in the perception vector are linearly transformed to obtain a key matrix and a value matrix; The temporal offset between the running state features and the business traffic features in the perception vector is encoded to obtain the temporal offset vector; The coupling strength between the query matrix and the corresponding time step in the key matrix is calculated based on the time offset vector, and then the attention weight is determined. The attention weights are applied to the corresponding business traffic features to obtain the fused features; The step of performing layer-by-layer iterative convolution processing on the fused features to obtain propagation features includes: According to the business communication path of the edge terminal in the terminal association graph, the fusion feature is subjected to layer-by-layer iterative convolution processing to obtain the propagation feature; The terminal association graph is constructed based on the business communication connection relationship of each power terminal in the power Internet of Things; the propagation weight between adjacent convolutional layers in the layer-by-layer iterative convolutional processing is determined based on the multi-dimensional registration attributes between adjacent nodes of the corresponding propagation segment in the terminal association graph; the multi-dimensional registration attributes include at least one of business affiliation relationship, access gateway relationship and transformer area affiliation relationship; The determination of the threat identification result for the edge terminal based on the fusion features, the temporal risk probability, and the propagation risk probability includes: The temporal risk and the propagation risk are weighted and fused to obtain the comprehensive risk value of the edge terminal; The fusion feature, the temporal risk probability, and the propagation risk probability are concatenated into vectors to form an enhanced feature; The enhanced features are nonlinearly mapped to obtain the risk category of the edge terminal; The risk category and overall risk value of the edge terminal are used as the threat identification result.
2. The method of claim 1, wherein, The formula for calculating the fusion feature is: in, As a feature of fusion, For the first p The running state characteristics of the first time step and the second time step q Attention weights among business traffic features at each time step For the first p The running state characteristics of the first time step and the second time step q The coupling strength between the business traffic characteristics at each time step For the first p The running state characteristics of the first time step and the second time step The coupling strength between the business traffic characteristics of each time step; , , These are the query matrix, key matrix, and value matrix, respectively. , , These are the weight matrices for the attention heads corresponding to the query, key, and value, respectively. , These are the running state feature vector and the service traffic feature vector, respectively, in the perception vector. , Each is a matrix , The Middle q Each time step corresponds to a vector. For matrix The Middle p Each time step corresponds to a vector. L This represents the current risk perception time window length. For the first p The running state characteristics of the first time step and the second time step q The temporal offset vector between the business traffic characteristics of each time step; The dimension of the key is T, and the superscript T denotes the transpose of the matrix.
3. The method according to claim 1, characterized in that, The formula for calculating the propagation characteristics is: in, , The first , Propagation characteristics of each propagation layer Nodes in the terminal association graph The set of neighboring nodes, The risk propagation weights between nodes i and j in the terminal association graph. Let be the trainable transformation matrix of the graph convolutional unit. The nonlinear activation function of the graph convolutional unit, , , These are the learnable weight parameters of the graph convolutional unit. This indicates whether nodes i and j in the terminal association graph belong to the same business application. This indicates whether the network access of nodes i and j in the terminal association diagram belongs to the same secure access gateway. This indicates whether nodes i and j in the terminal association graph are in the same power grid area. , , The values are all 0 or 1.
4. The method according to claim 1, characterized in that, After receiving the edge threat score, the following is also included: If the edge threat score is determined to be in the low-risk range, the operating status characteristics and edge threat score of the edge terminal are transferred to the cold data storage area.
5. The method according to claim 1, characterized in that, The characteristics of the service traffic include: Network quintuple information, service traffic statistical characteristics, and service attribute characteristics; The operational status characteristics include at least one of the following: CPU utilization, memory utilization, unauthorized external connections, unauthorized process execution, high-risk port opening, unauthorized peripheral device access, and critical configuration changes.
6. A cloud-based main station, characterized in that, include: The main station traffic mirroring probe is used to receive the identifier mapping relationship, the operating status characteristics and edge threat score of the edge terminal reported by the edge terminal; The identifier mapping relationship is used to indicate the association between the edge terminal and each service terminal accessing the edge terminal; it is also used to analyze the monitored service traffic and extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic when it is determined that the edge threat score of an edge terminal is within the medium risk threshold range. Based on the identification mapping relationship, the service traffic characteristics of each service terminal are associated with the operating status characteristics of the edge terminal associated with the service terminal identification to construct a perception vector; The terminal anomaly risk deep analysis engine is used to input the perception vector into a temporal topology fusion constraint network for threat analysis, and obtain the threat identification result of the edge terminal. The temporal topology fusion constraint network includes a multi-head attention unit, a graph convolution unit, a first classification unit, and a second classification unit; The multi-head attention unit is used to perform multi-head attention fusion on the perception vector to obtain fused features; The graph convolutional unit is used to perform layer-by-layer iterative convolution processing on the fused features to obtain the propagation features; The first classification unit is used to perform nonlinear activation on the fusion feature and the propagation feature respectively to obtain the temporal risk probability and the propagation risk probability; The second classification unit is used to determine the threat identification result of the edge terminal based on the fusion features, the temporal risk probability, and the propagation risk probability; The multi-head attention fusion is a feature fusion based on the temporal offset between the running state features and the business traffic features in the perception vector; The multi-head attention unit includes a linear transformation layer, a temporal coding layer, and an attention fusion layer; The linear transformation layer is used to: perform a linear transformation on the running state features in the perception vector to generate a query matrix; and also to perform a linear transformation on the business traffic features in the perception vector to obtain a key matrix and a value matrix. The temporal coding layer is used to: encode the temporal offset between the running state features and the business traffic features in the perception vector to obtain a temporal offset vector; The attention fusion layer is used to: calculate the coupling strength between the query matrix and the corresponding time step in the key matrix based on the temporal offset vector, and then determine the attention weights; The attention fusion layer is also used to: apply the attention weights to the corresponding business traffic features to obtain fused features; The graph convolution unit is specifically used for: According to the business communication path of the edge terminal in the terminal association graph, the fusion feature is subjected to layer-by-layer iterative convolution processing to obtain the propagation feature; The terminal association graph is constructed based on the business communication connection relationship of each power terminal in the power Internet of Things; the propagation weight between adjacent convolutional layers in the layer-by-layer iterative convolutional processing is determined based on the multi-dimensional registration attributes between adjacent nodes of the corresponding propagation segment in the terminal association graph; the multi-dimensional registration attributes include at least one of business affiliation relationship, access gateway relationship and transformer area affiliation relationship; The temporal topology fusion constraint network also includes a fusion unit; The fusion unit is used to: perform weighted fusion of the temporal risk and the propagation risk to obtain the comprehensive risk value of the edge terminal; The second classification unit is specifically used to: concatenate the fusion feature, the temporal risk probability, and the propagation risk probability into vectors to form an enhanced feature; input the enhanced feature into the classification unit for nonlinear mapping and output the risk category of the edge terminal; and use the risk category and comprehensive risk value of the edge terminal as the threat identification result.
7. The cloud-based main station according to claim 6, characterized in that, The formula for calculating the fusion feature is: in, As a feature of fusion, For the first p The running state characteristics of the first time step and the second time step q Attention weights among business traffic features at each time step For the first p The running state characteristics of the first time step and the second time step q The coupling strength between the business traffic characteristics at each time step For the first p The running state characteristics of the first time step and the second time step The coupling strength between the business traffic characteristics of each time step; , , These are the query matrix, key matrix, and value matrix, respectively. , , These are the weight matrices for the attention heads corresponding to the query, key, and value, respectively. , These are the running state feature vector and the service traffic feature vector, respectively, in the perception vector. , Each is a matrix , The Middle q Each time step corresponds to a vector. For matrix The Middle p Each time step corresponds to a vector. L This represents the current risk perception time window length. For the first p The running state characteristics of the first time step and the second time step q The temporal offset vector between the business traffic characteristics of each time step; The dimension of the key is T, and the superscript T denotes the transpose of the matrix.
8. The cloud-based main station according to claim 6, characterized in that, The formula for calculating the propagation characteristics is: in, , The first , Propagation characteristics of each propagation layer Nodes in the terminal association graph The set of neighboring nodes, The risk propagation weights between nodes i and j in the terminal association graph. Let be the trainable transformation matrix of the graph convolutional unit. The nonlinear activation function of the graph convolutional unit, , , These are the learnable weight parameters of the graph convolutional unit. This indicates whether nodes i and j in the terminal association graph belong to the same business application. This indicates whether the network access of nodes i and j in the terminal association diagram belongs to the same secure access gateway. This indicates whether nodes i and j in the terminal association graph are in the same power grid area. , , The values are all 0 or 1.
9. The cloud-based main station according to claim 6, characterized in that, The terminal anomaly risk deep analysis engine is also used for: If the edge threat score is determined to be in the low-risk range, the operating status characteristics and edge threat score of the edge terminal are transferred to the cold data storage area.
10. The cloud-based main station according to claim 6, characterized in that, The characteristics of the service traffic include: Network quintuple information, service traffic statistical characteristics, and service attribute characteristics; The operational status characteristics include at least one of the following: CPU utilization, memory utilization, unauthorized external connections, unauthorized process execution, high-risk port opening, unauthorized peripheral device access, and critical configuration changes.
11. A cloud-edge collaborative method for cross-domain threat perception of power terminals, characterized in that, Applied to edge terminals, including: The monitored service traffic is analyzed to extract the service terminal identifiers of each service terminal accessing the edge terminal, and the service terminal identifiers are associated with the edge terminal to generate an identifier mapping relationship; The edge terminal collects its own operational status characteristics and calculates its edge threat score based on the degree of deviation of the operational status characteristics from the baseline. The identification mapping relationship, the operating status characteristics, and the edge threat score are reported to the cloud master station, so that the cloud master station can analyze the monitored service traffic, extract the service traffic characteristics of each service terminal and the service terminal identification carried in the service traffic, and associate the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identification based on the identification mapping relationship to form a perception vector, and input the perception vector into the temporal topology fusion constraint network for threat analysis; The step of inputting the perception vector into a temporal topology fusion constraint network for threat analysis includes: Multi-head attention fusion is performed on the perception vectors to obtain fused features; The fused features are subjected to iterative convolutional processing layer by layer to obtain the propagation features; By applying nonlinear activation to the fusion feature and the propagation feature respectively, the temporal risk probability and the propagation risk probability are obtained; Based on the fusion features, the temporal risk probability, and the propagation risk probability, the threat identification result of the edge terminal is determined; The multi-head attention fusion is a feature fusion based on the temporal offset between the running state features and the business traffic features in the perception vector; The process of performing multi-head attention fusion on the perception vector to obtain fused features includes: The running status features in the perception vector are linearly transformed to generate a query matrix; the business traffic features in the perception vector are linearly transformed to obtain a key matrix and a value matrix; The temporal offset between the running state features and the business traffic features in the perception vector is encoded to obtain the temporal offset vector; The coupling strength between the query matrix and the corresponding time step in the key matrix is calculated based on the time offset vector, and then the attention weight is determined. The attention weights are applied to the corresponding business traffic features to obtain the fused features; The step of performing layer-by-layer iterative convolution processing on the fused features to obtain propagation features includes: According to the business communication path of the edge terminal in the terminal association graph, the fusion feature is subjected to layer-by-layer iterative convolution processing to obtain the propagation feature; The terminal association graph is constructed based on the business communication connection relationship of each power terminal in the power Internet of Things; the propagation weight between adjacent convolutional layers in the layer-by-layer iterative convolutional processing is determined based on the multi-dimensional registration attributes between adjacent nodes of the corresponding propagation segment in the terminal association graph; the multi-dimensional registration attributes include at least one of business affiliation relationship, access gateway relationship and transformer area affiliation relationship; The determination of the threat identification result for the edge terminal based on the fusion features, the temporal risk probability, and the propagation risk probability includes: The temporal risk and the propagation risk are weighted and fused to obtain the comprehensive risk value of the edge terminal; The fusion feature, the temporal risk probability, and the propagation risk probability are concatenated into vectors to form an enhanced feature; The enhanced features are nonlinearly mapped to obtain the risk category of the edge terminal; The risk category and overall risk value of the edge terminal are used as the threat identification result.
12. The method according to claim 11, characterized in that, The calculation of the edge threat score of the edge terminal based on the degree of deviation of the operating status characteristics from the baseline includes: For each operational indicator in the operational status characteristics, calculate the deviation of the current measured value from the corresponding baseline, and determine the abnormal score of the operational indicator based on the deviation. The abnormal scores of each operational indicator within the operational status feature are weighted and summed to obtain a comprehensive abnormal value; The comprehensive outlier values are numerically normalized to obtain the edge threat score of the edge terminal; The baseline for each operational indicator is determined based on the statistical measures of its historical values.
13. The method according to claim 11, characterized in that, After calculating the edge threat score, the following is also included: If the edge threat score is determined to be in the high-risk range, local security measures will be taken for the corresponding edge terminal.
14. The method according to claim 11, characterized in that, The process of constructing the identifier mapping relationship includes: The network interface card (NIC) service traffic of the edge terminal is obtained by port mirroring, and the service terminal identifier in the NIC service traffic is extracted by parsing the MQTT northbound service traffic protocol. The business terminal identifier is bound to the edge terminal identifier of the edge terminal itself to form the identifier mapping relationship.
15. An edge terminal, characterized in that, Includes a terminal security monitoring probe, wherein the terminal security monitoring probe includes: The mapping generation unit is used to parse the monitored service traffic, extract the service terminal identifiers of each service terminal accessing the edge terminal, and associate the service terminal identifiers with the edge terminal to generate an identifier mapping relationship. The threat calculation unit is used to collect its own operational status characteristics and calculate the edge threat score of the edge terminal based on the degree of deviation of the operational status characteristics from the baseline. The reporting unit is used to report the identifier mapping relationship, the operating status characteristics, and the edge threat score to the cloud master station, so that the cloud master station can analyze the monitored service traffic, extract the service traffic characteristics of each service terminal and the service terminal identifier carried in the service traffic, and associate the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identifier based on the identifier mapping relationship to form a perception vector, and input the perception vector into the temporal topology fusion constraint network for threat analysis; The step of inputting the perception vector into a temporal topology fusion constraint network for threat analysis includes: Multi-head attention fusion is performed on the perception vectors to obtain fused features; The fused features are subjected to iterative convolutional processing layer by layer to obtain the propagation features; By applying nonlinear activation to the fusion feature and the propagation feature respectively, the temporal risk probability and the propagation risk probability are obtained; Based on the fusion features, the temporal risk probability, and the propagation risk probability, the threat identification result of the edge terminal is determined; The multi-head attention fusion is a feature fusion based on the temporal offset between the running state features and the business traffic features in the perception vector; The process of performing multi-head attention fusion on the perception vector to obtain fused features includes: The running status features in the perception vector are linearly transformed to generate a query matrix; the business traffic features in the perception vector are linearly transformed to obtain a key matrix and a value matrix; The temporal offset between the running state features and the business traffic features in the perception vector is encoded to obtain the temporal offset vector; The coupling strength between the query matrix and the corresponding time step in the key matrix is calculated based on the time offset vector, and then the attention weight is determined. The attention weights are applied to the corresponding business traffic features to obtain the fused features; The step of performing layer-by-layer iterative convolution processing on the fused features to obtain propagation features includes: According to the business communication path of the edge terminal in the terminal association graph, the fusion feature is subjected to layer-by-layer iterative convolution processing to obtain the propagation feature; The terminal association graph is constructed based on the business communication connection relationship of each power terminal in the power Internet of Things; the propagation weight between adjacent convolutional layers in the layer-by-layer iterative convolutional processing is determined based on the multi-dimensional registration attributes between adjacent nodes of the corresponding propagation segment in the terminal association graph; the multi-dimensional registration attributes include at least one of business affiliation relationship, access gateway relationship and transformer area affiliation relationship; The determination of the threat identification result for the edge terminal based on the fusion features, the temporal risk probability, and the propagation risk probability includes: The temporal risk and the propagation risk are weighted and fused to obtain the comprehensive risk value of the edge terminal; The fusion feature, the temporal risk probability, and the propagation risk probability are concatenated into vectors to form an enhanced feature; The enhanced features are nonlinearly mapped to obtain the risk category of the edge terminal; The risk category and overall risk value of the edge terminal are used as the threat identification result.
16. The edge terminal according to claim 15, characterized in that, The threat calculation unit is specifically used for: For each operational indicator in the operational status characteristics, calculate the deviation of the current measured value from the corresponding baseline, and determine the abnormal score of the operational indicator based on the deviation. The abnormal scores of each operational indicator within the operational status feature are weighted and summed to obtain a comprehensive abnormal value; The comprehensive outlier values are numerically normalized to obtain the edge threat score of the edge terminal; The baseline for each operational indicator is determined based on the statistical measures of its historical values.
17. The edge terminal according to claim 15, characterized in that, The terminal security monitoring probe also includes: The security handling unit is used to perform local security handling on the corresponding edge terminal if the edge threat score is determined to be in the high-risk range.
18. The edge terminal according to claim 15, characterized in that, The mapping generation unit is specifically used for: The network interface card (NIC) service traffic of the edge terminal is obtained by port mirroring, and the service terminal identifier in the NIC service traffic is extracted by parsing the MQTT northbound service traffic protocol. The business terminal identifier is bound to the edge terminal identifier of the edge terminal itself to form the identifier mapping relationship.
19. A cloud-edge collaborative method for cross-domain threat perception of power terminals, characterized in that, include: The edge terminal parses the monitored service traffic, extracts the service terminal identifiers of each service terminal accessing the edge terminal, and associates the service terminal identifiers with the edge terminal to generate an identifier mapping relationship; The edge terminal collects its own operating status characteristics and calculates the edge threat score based on the degree of deviation of the operating status characteristics from the baseline; the identification mapping relationship, the operating status characteristics, and the edge threat score are reported to the cloud master station. When the cloud master station determines that the edge threat score of an edge terminal is within the medium risk threshold range, it analyzes the monitored business traffic and extracts the business traffic characteristics of each business terminal and the business terminal identifier carried in the business traffic. Based on the identifier mapping relationship, the cloud master station associates the service traffic characteristics of each service terminal with the operating status characteristics of the edge terminal associated with the service terminal identifier to construct a perception vector; and inputs the perception vector into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal. The step of inputting the perception vector into a temporal topology fusion constraint network for threat analysis to obtain the threat identification result of the edge terminal includes: Multi-head attention fusion is performed on the perception vectors to obtain fused features; The fused features are subjected to iterative convolutional processing layer by layer to obtain the propagation features; By applying nonlinear activation to the fusion feature and the propagation feature respectively, the temporal risk probability and the propagation risk probability are obtained; Based on the fusion features, the temporal risk probability, and the propagation risk probability, the threat identification result of the edge terminal is determined; The multi-head attention fusion is a feature fusion based on the temporal offset between the running state features and the business traffic features in the perception vector; The process of performing multi-head attention fusion on the perception vector to obtain fused features includes: The running status features in the perception vector are linearly transformed to generate a query matrix; the business traffic features in the perception vector are linearly transformed to obtain a key matrix and a value matrix; The temporal offset between the running state features and the business traffic features in the perception vector is encoded to obtain the temporal offset vector; The coupling strength between the query matrix and the corresponding time step in the key matrix is calculated based on the time offset vector, and then the attention weight is determined. The attention weights are applied to the corresponding business traffic features to obtain the fused features; The step of performing layer-by-layer iterative convolution processing on the fused features to obtain propagation features includes: According to the business communication path of the edge terminal in the terminal association graph, the fusion feature is subjected to layer-by-layer iterative convolution processing to obtain the propagation feature; The terminal association graph is constructed based on the business communication connection relationship of each power terminal in the power Internet of Things; the propagation weight between adjacent convolutional layers in the layer-by-layer iterative convolutional processing is determined based on the multi-dimensional registration attributes between adjacent nodes of the corresponding propagation segment in the terminal association graph; the multi-dimensional registration attributes include at least one of business affiliation relationship, access gateway relationship and transformer area affiliation relationship; The determination of the threat identification result for the edge terminal based on the fusion features, the temporal risk probability, and the propagation risk probability includes: The temporal risk and the propagation risk are weighted and fused to obtain the comprehensive risk value of the edge terminal; The fusion feature, the temporal risk probability, and the propagation risk probability are concatenated into vectors to form an enhanced feature; The enhanced features are nonlinearly mapped to obtain the risk category of the edge terminal; The risk category and overall risk value of the edge terminal are used as the threat identification result.
20. A cloud-edge collaborative cross-domain threat perception system for power terminals, characterized in that, include: The cloud-based main station as described in any one of claims 6-10 and / or the edge terminal as described in any one of claims 15-18.
21. A readable storage medium, characterized in that, It contains an executable program, which, when executed, implements the method as described in any one of claims 1-5 and 11-14.
Citation Information
Patent Citations
Multi-stage distributed monitoring and anti-permeation system for power terminal
CN112261041A
Internet of Things terminal safety monitoring protection method and system
CN113569236A
Power system protocol threat hunting method and system based on cloud edge collaboration and TTP model
CN121690643A
Cloud-edge collaborative industrial scene risk discovery and evaluation method, system and related device
CN121690667A
