Distributed behavior authentication method oriented to computing power network

Abstract

The invention provides a distributed behavior authentication method oriented to a computing power network. The distributed behavior authentication method comprises the following steps: step 1, constructing a feature extractor; the terminals of the computing power network cooperatively train an identity feature extractor, and the identity feature extractor is applied to each terminal to generate identity credential representation vectors of the terminals; 2, low-delay sensing deployment is carried out; the regional cloud server performs similarity calculation on different identity certificates of all the terminals, deploys authentication services to an edge authentication server according to needs, and minimizes the delay time of terminal mobile authentication; 3, performing multi-stage decision verification; by applying a hierarchical clustering algorithm, each edge server creates a corresponding identity template for each terminal according to the received information, the identity template is composed of a unique identity label and a corresponding identity certificate, and a comprehensive decision is made by using the similarity of behavior modes. The method provided by the invention meets the privacy requirement of the user, and has the characteristics of high authentication performance, low overhead and high security.

Description

Technical Field

[0001] The present invention relates to the technical field of authentication systems, and in particular to a distributed behavior authentication method for computing power networks. Background Art

[0002] Computing Power Networks (CPNs) are a new type of information infrastructure that flexibly allocates computing, storage, and network resources across the cloud, edge, and end-users based on demand. Their core concept is to provide computing resource services, enabling users to dynamically schedule computing tasks to appropriate locations based on business needs. In this new network, businesses and individual users require not only network and cloud services but also the flexibility to manage the location of computing tasks. Similar to how telephone networks improve communication efficiency and the internet enhances collaboration, CPNs aim to enhance the collaborative efficiency of cloud, edge, and end-user computing, improve the user experience, and reduce latency.

[0003] Behavioral authentication is an authentication method that verifies a user's identity by analyzing and identifying their behavioral patterns and characteristics. This authentication method relies on the unique behaviors people exhibit when using devices and systems, such as keyboard input habits, mouse movement patterns, pressure levels, and consumption logs, rather than relying on traditional passwords or biometrics. By continuously monitoring and analyzing user behavior, behavioral authentication systems can establish a baseline for user behavior and trigger the authentication process if abnormal behavior or discrepancies with the baseline are identified. This approach helps improve security, reduces reliance on cumbersome passwords, and allows authentication to occur in the background, making the user experience more seamless and convenient.

[0004] In a computing network, centralized behavior authentication typically involves the following steps: 1. User data collection: First, the system collects behavioral data from different terminal users. This data may include the user's keyboard input method, mouse movement pattern, login time, login location and other information; the collected information is stored in a centralized server group.

[0005] 2. Model building: After collecting data, the system uses machine learning and other technologies to build a user behavior model. By analyzing user behavior patterns, the system can identify user characteristics and habits.

[0006] 3. Real-time monitoring: Once a user attempts to log in to the system, the centralized behavioral authentication system monitors the user's behavior in real time. The system compares the user's current behavior with the previously established model to determine whether the user is a genuine user.

[0007] Existing centralized behavior authentication technologies have some shortcomings, especially when facing emerging technologies and demands, which may lead to the following problems: Privacy issues: Centralized behavior authentication requires collecting a large amount of user behavior data, which may involve user privacy issues. In particular, data stored on centralized servers is easily targeted by hackers. Once the centralized server data is leaked, serious consequences will occur. Accuracy of behavioral authentication: The existing model building process may not accurately depict user behavior patterns, especially when there is insufficient behavioral data for authentication, which may affect the accuracy of authentication. Network delay: Existing methods involve a large amount of user data flow, which will result in additional loan overhead and bring additional delay.

[0008] Single point of failure: Centralized authentication systems have the risk of single point of failure. Once a server fails or is attacked, the entire authentication system may be paralyzed, resulting in service unavailability.

[0009] Existing behavioral authentication methods rely on a fully trusted central server to collect data. This approach is oriented towards the distributed environment of new computing networks. The collection of behavioral data faces heavy communication overhead, making it difficult to ensure that important business-sensitive data must be processed and stored locally. Summary of the Invention

[0010] In response to the problems existing in the existing technology, the present invention needs to further reduce the authentication delay while complying with data management specifications, and establish an authentication system with priority on authentication scalability to meet the cross-domain low-latency and high-precision requirements of computing power network identity authentication.

[0011] Technical solution of the present invention: A distributed behavior authentication method for a computing power network includes the following steps: Step 1: Feature extractor construction The terminals of the computing power network jointly train a deep neural network model. After the model training is completed, the hidden parameter model before its last fully connected layer is used to generate the identity embedding vector, which is called the identity feature extractor using hidden layer parameters.

[0012] After the feature extractor is established, it is applied to each terminal to generate the terminal's identity credential representation vector; Step 2: Low-latency awareness deployment After the regional cloud server completes the collection of identity credential representation vectors of all terminals, it calculates the similarity of these different identity credentials, and then deploys the authentication service to the edge authentication server on demand, minimizing the delay time of terminal mobile authentication.

[0013] Step 3: Multi-stage decision verification By applying the hierarchical clustering algorithm, each edge server creates a corresponding identity template for each terminal based on the received information, which consists of a unique identity identifier and corresponding identity credentials, and makes comprehensive decisions based on the similarity of behavior patterns.

[0014] Beneficial effects 1. Protect user privacy: The solution should meet user privacy requirements and consider information security. The process of establishing the identity verification system should ensure that behavioral data is not shared or leaked.

[0015] 2. High authentication performance: The solution should provide users with highly accurate identity authentication services and prevent malicious users from gaining access to the system.

[0016] 3. Low Overhead: Considering the strong real-time requirements of edge computing services, authentication should not require additional operations on the device to avoid introducing additional delays and inconvenience. In addition, the communication process has low overhead, as shown in Table 3. 4. Strong security: The scheme should be able to resist various attacks, including replay attacks, model inversion attacks, and impersonation attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of the distributed behavior authentication method for computing power network of the present invention; Figure 2 This is the technical roadmap for the distributed behavior authentication method for computing power networks of the present invention; Figure 3 Schematic diagram of the safety test results of an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The technical solution provided by this application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of this application will become more apparent with reference to the following description.

[0019] In this embodiment, the computing power network includes the following nodes: a regional cloud server, several terminals, and several edge authentication servers.

[0020] in, Regional cloud servers: These servers provide cloud computing resources, typically deployed in data centers within a geographic region. Within a computing network, regional cloud servers are responsible for generating authentication credentials and coordinating distributed authentication.

[0021] Terminal: Refers to a device connected to the network. In a computing network, a terminal is the interface through which users interact with the computing network, initiating authentication requests and receiving authentication results. The terminal is also responsible for collecting and storing user behavior data. Specifically, a terminal can be a mobile device used by a user.

[0022] Edge authentication server: Located at the edge of the network, the edge authentication server is responsible for processing authentication requests initiated by terminal devices. Its role is to quickly respond to authentication requests at the edge, reducing the burden on cloud servers in the central area and improving authentication response speed.

[0023] The base station is located between the terminal and the edge authentication server, and is responsible for data transmission, request forwarding and load balancing during the terminal authentication process.

[0024] A distributed behavior authentication method for computing power network, comprising the following steps: (e.g. Figure 1 ) Step 1: Feature extractor construction The terminals of the computing power network jointly train a deep neural network model. After the model training is completed, the hidden parameter model before its last fully connected layer is used to generate the identity embedding vector, which is called the identity feature extractor using hidden layer parameters.

[0025] After the feature extractor is established, it is applied to each terminal to generate the terminal's identity credential representation vector; Step 2: Low-latency awareness deployment After the regional cloud server completes the collection of identity credential representation vectors of all terminals, it calculates the similarity of these different identity credentials, and then deploys the authentication service to the edge authentication server on demand, minimizing the delay time of terminal mobile authentication.

[0026] Step 3: Multi-stage decision verification By applying the hierarchical clustering algorithm, each edge server creates a corresponding identity template for each terminal based on the received information, which consists of a unique identity identifier and corresponding identity credentials, and makes comprehensive decisions based on the similarity of behavior patterns.

[0027] The specific process is as follows: Step 1, feature extractor construction: Unlike the traditional paradigm of centralized data collection based on behavioral identity authentication, in real-world scenarios, data is diverse and scattered across various locations. In this scenario, data comes from different sources and has different owners. Data owners cannot directly share their private behavioral data, making it impossible to build a large-scale centralized data platform. Furthermore, it is impossible to rely on a model trained on centralized data to generate identity credentials for different users.

[0028] To achieve distributed terminal behavior authentication, the regional cloud server and the terminal first collaboratively build a deep neural network model and use federated learning methods to collaboratively train the model without leaking private behavior data. After the model training is completed, the hidden parameter model before the last fully connected layer is the identity feature extractor. The identity feature extractor is applied to each terminal to generate the terminal's behavior representation credentials, that is, each user's independent identity credentials.

[0029] a. Auxiliary task establishment: First, different terminals jointly train a deep neural network model (DNN). After the model training is completed, it will not be used to process new authentication requests. What is needed are some parameters generated after the model training is completed, that is, the hidden parameter model before the last fully connected layer to generate the identity embedding vector. Therefore, it is called an identity feature extractor using hidden layer parameters, and the task used to train the model is also called an auxiliary task.

[0030] The deep neural network model training data is the terminal's behavioral data, which is collected by the terminal device through sensors or other data sources, such as tapping patterns, sliding methods, pressing strength, payment records, etc. Each user's identity information will be associated with it as a label. The label can be a unique user ID, which is used to indicate the user identity to which each data point belongs.

[0031] As an example, the deep neural network architecture includes two fully connected layers and a LeakyReLU activation function followed by a Softmax function. Dropout is applied to the network input with a probability of 0.5. The gradient clipping threshold δ is set to 0.15, and the Laplace noise intensity λ is set to 0.1. The optimizer uses the Adam algorithm with an initial learning rate of 0.05 and a weight decay coefficient of 0.0002.

[0032] The auxiliary task is a multi-classification task used to embed the learning behavior data of the feature extractor during training.

[0033] Specifically, for each terminal, the regional cloud will first detect the status of the terminal. If the deep neural network model does not exist, the terminal will create a deep neural network model; otherwise, if the model has been created, it will update the model.

[0034] b. Collaborative model training: During the training process, in order to ensure that different terminal devices can collaboratively train the model without leaking private behavior data, the federated learning method is used to share parameters during the training process, thus ensuring that the different terminal behavior representations generated by the subsequent hidden layer are measurable.

[0035] No. Behavior data collection of a terminal Can be divided into two subsets with no common elements and satisfying: in and Respectively represent The data set of normal behavior and abnormal behavior of each terminal.

[0036] The local model of each terminal has the same network architecture and initialization parameters, and the model parameters are updated through backpropagation. is the collection of all terminals, yes The number of samples in , and Represents the The gradient of each user, if the learning rate is defined as , when the global parameter is When the round is updated, each terminal executes in parallel: in, For the Aggregate weight of the round, send new parameters after completing the update To the regional cloud server, and then the regional cloud server performs the operation: Perform weighted averaging on the weights from different terminals, where K represents the number of all users; then return the calculated parameters To different terminals. After receiving the new model parameters, the terminal updates the local model.

[0037] pass To express the model parameters Next sample Finally, the algorithm defines the following objective function: in, is the prediction error, Represents the number of local samples of the kth client.

[0038] After the model training is completed, the hidden parameter model before the last fully connected layer is the identity feature extractor.

[0039] c. Terminal identity credential vectorization: After the identity feature extractor is established, it is applied to each terminal to generate the terminal's behavior representation credentials. Specifically, when the input data , after being processed by the feature extractor, the representation vector is recorded as .

[0040] Characterization set of terminal behavior data as follows: in and Respectively represent The characterization set of normal behavior data and abnormal behavior data of each terminal.

[0041] Furthermore, the terminal's identity credential representation vector can be obtained ,as follows: in yes The number of samples in the dataset is counted and sent to the regional cloud server.

[0042] Step 2: Low-latency awareness deployment: When the regional cloud receives identity credential representation vectors from different terminals, it will calculate the similarity of these different identity credentials and, considering the requirements of latency and load balancing, deploy the authentication service to the optimal edge location on demand, such as Figure 2 shown.

[0043] Because the deployment location of edge servers is crucial to the latency of mobile authentication, especially in multi-terminal network computing environments, using only cloud servers in the same region will result in long authentication delays. Therefore, authentication methods in new network computing models cannot be separated from edge cloud collaboration. Furthermore, the location of edge servers can significantly impact authentication efficiency. Improper edge server deployment can easily lead to overloading some edge servers while underutilizing others. Existing solutions rarely consider the importance of proper edge server deployment for authentication.

[0044] The low-latency awareness deployment specifically includes the following processes: a. The regional cloud server generates an identity metric matrix and clusters these credentials into different decision clusters: First, during the terminal registration phase, the identity credentials of all terminals received by the regional cloud are represented to generate an identity measurement matrix, which is expressed as: Among them, The columns represent The identity credential representation vector of each terminal, ;No. The row represents the identity certificate Dimensions ， ; Represents the identity vector No. Dimensional features.

[0045] Furthermore, based on this matrix, these credentials are clustered into different decision clusters by applying the hierarchical clustering algorithm. The different decision cluster sets obtained are recorded as .

[0046] b. Edge latency deployment optimization and load balancing: The set of edge authentication servers in the network is defined as the following set: ,in Represents a specific edge authentication server and assumes that the resource processing capabilities are the same during different authentication service periods.

[0047] The set of all base station resources that need to be connected during the terminal authentication process is defined as: ,in Represents a specific base station. is the total number of base stations.

[0048] Edge servers are responsible for forwarding terminal authentication requests from base stations. Assuming the network bandwidth can transmit the same amount of data per unit time, the latency of an edge server can be reduced to the sum of the latency of the base stations it communicates with. The core issue of the interaction between terminal authentication and edge servers can be categorized as the deployment location of the edge authentication server and the on-demand allocation of request forwarding base stations. For edge server deployment, the latency optimization goal can be as follows: in, Represents the current A feasible edge deployment solution, represents the total delay of authentication, Represents the current authentication server All base stations that can be covered, Representative base station To the edge authentication server Delays; Furthermore, the load balancing metrics under this deployment scheme can be expressed as: in Represents the edge authentication server The total load of the covered base stations is calculated as follows: , Representative base station Similarly, Represents the edge authentication server Total load of covered base stations; Represents the edge server and The difference between the loads; therefore, we can get a weighted objective optimization function that comprehensively considers the deployment location of the authentication server and the load balancing of the distribution scheme: in, is the weight. And it is necessary to satisfy that each base station covers the only edge server for communication, and each base station has a corresponding edge server for communication, that is , ; At the same time, the load of the edge authentication server is equal to the sum of the loads of the base station, that is, .

[0049] Step 3, multi-stage decision verification: By applying the hierarchical clustering algorithm, each edge server creates a corresponding identity template for each terminal based on the received information, which consists of a unique identity identifier and corresponding identity credentials, and makes comprehensive decisions based on the similarity of behavior patterns.

[0050] a. Edge topology map construction: For terminals with similar patterns, create corresponding identity maps for them and build a topology map based on the results. The node set is recorded as ,in Represents the terminals in the cluster, } ,in Represents an edge between two terminals.

[0051] The scale of identity credential representation generation for all terminals in the same decision cluster is The identity metric matrix of in, represents the dimension of the identity credential vector, Represents the number of registered terminals in the same cluster; Connection Credentials Vector and The edge between Represents an edge The weight (distance) is: (1) in and Represents vectors and of Dimensional identity characteristics.

[0052] Furthermore, the edge server sorts the weights within the cluster from high to low, expressed as , here Represents the number of terminals in the same decision cluster, is the terminal weight that is closest to a specific authentication terminal mode, is the terminal weight that deviates most from a specific authentication terminal pattern. These credential vectors are stored in the corresponding edge server for further authentication.

[0053] b. Collaborative verification between individuals and collective intelligence: In the verification phase, A terminal generates new behavior data. For the privacy-sensitive data of the terminal, the input data is trained and generated by the feature extractor, and the generated vector representation is recorded as , and sends it to the nearest edge server for verification. When the corresponding edge server receives the corresponding authentication request, it will make a two-stage authentication decision based on the collaborative authentication committee mechanism. That is, it uses the similarity of behavioral patterns to make a comprehensive decision, specifically including the individual authentication decision and the group authentication decision.

[0054] Individual certification decisions: Specifically, during the first phase of authentication, the input vector is measured The behavior template stored in the terminal The result of this stage is called the decision of individual authentication. For the input data, if the distance difference is too large compared with the stored behavior pattern, it may be a request initiated by an attacker. Therefore, the distance between the input behavior data and the behavior template is measured by the change of the intensity relationship. According to the above distance formula (1), , for The percentage change in the intensity relationship is further calculated according to the following formula : in Represents the weight of one's own identity vector, if A positive value indicates that the behaviors are closer, and a negative value indicates that they are farther apart.

[0055] Qunzhi Certification’s decision: In the second phase of verification, the input vector is measured The change in the strength relationship compared with other terminals in the original topology, so this process is called group-level verification. Specifically, the generated behavior representation vector is transmitted to the edge server, and a new matrix copy is created, replacing the original identity embedding vector, and the existing identity credential vector is recalculated with other terminals in the same decision cluster according to the previous distance formula. The change in the intensity relationship between Here, the change in strength relationship is calculated based on cosine similarity, so it better captures how similar two vectors are in direction.

[0056] The edge server stores the input vector and the weights of other users in the same decision cluster, denoted as For users waiting for authentication , firstly, the connection strength is adjusted by SoftMax Normalize and get the corresponding weights , and thus calculate the change in the percentage of connection strength within the same decision cluster according to the following formula: like The increase in intensity indicates that the new behavior vector tends to shorten the distance between the behavior pattern to be authenticated and the users with similar behavior patterns in the same decision cluster; otherwise, if This decrease indicates that the new behavior vector tends to shift the distance between the behavior pattern to be authenticated and the users with similar behavior patterns in the same decision cluster. This also indicates that the behavior pattern is likely initiated by an attacker.

[0057] c. Adaptive decision aggregation: Design an adaptive two-stage verification weight aggregation method; The terminal to be authenticated, if the distribution of its characterization vector is recorded as , the corresponding variance is recorded as ; The distribution within a decision cluster is recorded as , the corresponding variance is recorded as .definition The adaptive parameters are To weight the two-stage authentication results, the calculation formula is as follows: therefore The weight of Therefore, the final authentication result combines the two-stage decision-making. In order to further adaptively integrate the results of individual decision-making and group decision-making, and considering the gap between the distributions, the final authentication result is fused according to the following formula: This process can be formally described as follows: If , we can get ;otherwise .in Represents the final authentication result returned by the edge-deployed authentication server. Represents the authentication threshold, with a value range of [0,1]. It is set to 0.8 in the payment scenario.

[0058] Comparative test: Comparative testing was conducted in a mobile payment scenario, comparing the proposed method (EBA) with FL, PFL, and LBA, using metrics such as F1-score, Recall, Precision, and Disturbance. EBA, which stands for Edge Behavior Authentication, is the proposed terminal edge behavior authentication solution. It achieves the best performance within a distributed authentication framework without the risk of privacy leakage.

[0059] We used a real-world dataset of online payment transactions obtained from a commercial bank. This dataset includes three consecutive months of B2C transaction records from 69,549 users who conducted online payment transactions via smartphone. To simplify the dataset, we removed features with sparsely significant values ​​from the original transaction set. Ultimately, we selected eight attributes for model construction. The meaning of these attributes is detailed in Table 1. The data underlying transaction behavior reflects unique shopping preferences, tastes, and spending habits, providing insights into user spending patterns, preferences, and responses to promotions and discounts.

[0060] Table 1 Selected attributes and descriptions of the transaction dataset The results are shown in Table 2.

[0061] Table 2 Comparison of test indicators of the method of the present invention As can be seen, the proposed EBA method achieves a high recall of 90.6% on an online payment transaction dataset. Compared to state-of-the-art authentication methods in a distributed device setting, EBA achieves a 14.4% improvement in precision on the online payment transaction dataset. Specifically, EBA ensures robust performance on the online payment transaction dataset, ensuring high recall (>90%) and low noise (<0.2%). Therefore, the proposed method can be considered a viable alternative to existing traditional authentication schemes in edge computing for smart mobile payment services with a frictionless authentication process.

[0062] Safety testing: A latent attack aims to mimic normal user behavior without being detected. The success probability of the attack is calculated for the extreme case where an elusive attacker attempts to control more than half of the seats within a specific decision cluster. Let p∈(0,1) denote the percentage of users in a specific decision cluster, and let q∈(0,1) denote the percentage of elusive attackers in the same decision cluster. Figure 3 Results depicting the probability of attack success for different values ​​of p and q are shown in Figure 1. It can be inferred that the proposed user authentication scheme is susceptible to a certain degree of risk only under specific conditions: when there is a high concentration of elusive attackers (high q) and a low participation rate of user decisions (low p) within a particular cluster. Even in these extreme cases, the identified risk remains below 0.1. Therefore, it can be asserted that this risk can be effectively mitigated by maintaining a high decision ratio p, as shown in the simulation results.

[0063] While the potential for elusive attackers in the real world is minimal, due to significant technical challenges and cost overhead, ensuring the scheme's security in extreme cases can be achieved by setting a reasonable value for p. Traditional authentication methods, which rely solely on individual user identity verification, are unable to protect against latent attacks. This also demonstrates the effectiveness of group-level authentication facilitated by the C3 mechanism in the EBA scheme. To enhance the robustness of the scheme, it is recommended to consider configuring a highly defensive edge server within the proposed framework.

[0064] Overhead test: Regarding communication overhead: During training, the parameters required to be transmitted in the EBA scheme were analyzed. In the proposed EBA scheme, only the DNN model parameters (weights and biases) need to be transmitted, and these are extremely lightweight. In contrast, traditional CBA methods require the transmission of the original dataset, significantly increasing communication traffic. As shown in Table 3, the information transmission in the online payment transaction dataset only consumes 3.5MB of communication traffic.

[0065] Table 3 Communication overhead of implementation cases Analysis of Time Cost: Verification latency primarily includes propagation delay, transmission delay, queuing delay, and processing delay. Processing delay includes local processing delay and verification server processing delay. Propagation delay, transmission delay, queuing delay, and verification server processing delay can be assumed constant across distributed authentication methods. The primary difference lies in local processing delay. Test results show that the proposed method generates local verification embeddings in online payment transaction data in 0.058ms, effectively handling frequent authentication requests in edge computing environments.

[0066] The above description is only a description of the preferred embodiments of the present application and does not limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical content should be regarded as equivalent valid embodiments and fall within the scope of protection of the technical solution of the present application.

Claims

1. A distributed behavior authentication method for computing power network, characterized in that: The following steps are involved: Step 1: Feature extractor construction The terminals of the computing network work together to train a deep neural network model. After the model is trained, the hidden parameter model before the last fully connected layer is used to generate the identity embedding vector, which is called the identity feature extractor using the hidden layer parameters. After the feature extractor is established, it is applied to each terminal to generate the terminal's identity credential representation vector; Step 2: Low-latency awareness deployment After the regional cloud server completes the collection of identity credential representation vectors for all terminals, it calculates the similarity of these different identity credentials and then deploys the authentication service to the edge authentication server as needed, minimizing the delay of terminal mobile authentication. Step 3: Multi-stage decision verification By applying the hierarchical clustering algorithm, each edge server creates a corresponding identity template for each terminal based on the received information, which consists of a unique identity identifier and corresponding identity credentials, and makes comprehensive decisions based on the similarity of behavior patterns.

2. A distributed behavior authentication method for computing power network according to claim 1, characterized in that: Step 1 is as follows: The regional cloud server and the terminal first collaboratively build a deep neural network model and use federated learning methods to collaboratively train the model without leaking private behavior data; After model training is complete, the hidden parameter model before the last fully connected layer becomes the identity feature extractor. This feature extractor is applied to each terminal to generate a behavior representation credential for the terminal, i.e., each user's independent identity credential. a. Auxiliary task establishment: First, different terminals collaboratively train a deep neural network model. After the model is trained, it will not be used to process new authentication requests. Instead, the hidden parameter model before the last fully connected layer is used to generate the identity embedding vector. The tasks used to train the model are also called auxiliary tasks; Each user's identity information will serve as a label; The auxiliary task is a multi-classification task, which is used to embed the learning behavior data of the feature extractor during training; b. Collaborative model training: During the training process, the federated learning method is used to share parameters during the training process; No. Behavior data collection of a terminal Can be divided into two subsets with no common elements and satisfying: in and Respectively represent The data set of normal behavior and abnormal behavior of each terminal; The local model of each terminal has the same network architecture and initialization parameters, and the model parameters are updated through back propagation; is the collection of all terminals, yes The number of samples in , and Represents the The gradient of the user, the learning rate is defined as , when the global parameter is When the round is updated, each terminal executes in parallel: in, For the Aggregate weight of the round, send new parameters after completing the update To the regional cloud server, and then the regional cloud server performs the operation: Perform weighted averaging on the weights from different terminals, where K represents the number of all users; then return the calculated parameters To different terminals; the terminal updates the local model after receiving the new model parameters; pass To express the model parameters Next sample Finally, the algorithm defines the following objective function: in, is the prediction error, Represents the number of local samples of the kth client; After the model training is completed, the hidden parameter model before the last fully connected layer is the identity feature extractor; c. Terminal identity credential vectorization: After the identity feature extractor is established, it is applied to each terminal to generate the terminal's behavior representation credentials; specifically, when the input data , after being processed by the feature extractor, the representation vector is recorded as ; Characterization set of terminal behavior data as follows: in and Respectively represent A representation set of normal behavior data and a representation set of abnormal behavior data of each terminal; Terminal identity credential representation vector ,as follows: in yes The number of samples in Send the terminal's identity credential representation vector to the regional cloud server.

3. A distributed behavior authentication method for computing power network as claimed in claim 2, characterized in that: The deep neural network model training data is the terminal's behavioral data, which is collected by the terminal device through sensors or other data sources, including: tapping patterns, sliding methods, pressing strength, and payment records.

4. A distributed behavior authentication method for computing power network according to claim 1, characterized in that: In step 2, after the regional cloud receives identity credential representation vectors from different terminals, it calculates the similarity of these different identity credentials and deploys the authentication service to the optimal edge location on demand, taking into account latency and load balancing requirements. The low-latency awareness deployment is specifically The following processes are included: a. The regional cloud server generates an identity metric matrix and clusters these credentials into different decision clusters: First, during the terminal registration phase, the identity credentials of all terminals received by the regional cloud are represented to generate an identity measurement matrix, which is expressed as: Among them, The columns represent The identity credential representation vector of each terminal, ;No. The row represents the identity certificate Dimensions ， ; Represents the identity vector No. dimensional features; Based on the identity metric matrix, these credentials are clustered into different decision clusters by applying the hierarchical clustering algorithm. The different decision cluster sets obtained are recorded as ; b. Edge latency deployment optimization and load balancing: The set of edge authentication servers in the network is defined as the following set: ,in Represents a specific edge authentication server, and assumes that the resource processing capacity is the same during different authentication service periods; The set of all base station resources that need to be connected during the terminal authentication process is defined as: ,in Represents a specific base station. is the total number of base stations; The edge server is responsible for handling terminal authentication requests forwarded by base stations. Assuming the network bandwidth can transmit the same amount of data per unit time, the edge server's latency can be converted into the sum of the latency of the base stations it communicates with. The core issue of the interaction between terminal authentication and edge servers can be converted into the deployment location of the edge authentication server and the on-demand allocation of request forwarding base stations. For edge server deployment, the latency optimization goals are as follows: in, Represents the current A feasible edge deployment solution, represents the total delay of authentication, Represents the current authentication server All base stations that can be covered, Representative base station To the edge authentication server Delays; The load balancing metrics under this deployment scheme are expressed as: in Represents the edge authentication server The total load of the covered base stations is calculated as follows: , Representative base station load; Represents the edge authentication server Total load of covered base stations; Represents the edge server and The difference between the loads; The weighted objective optimization function for load balancing that comprehensively considers the authentication server deployment location and distribution scheme is: in, is the weight, and it must satisfy that the edge server covering each base station is unique, and each base station has a corresponding edge server for communication, that is, , ; At the same time, the load of the edge authentication server is equal to the sum of the loads of the base station, that is, .

5. A distributed behavior authentication method for computing power network according to claim 1, characterized in that: Specifically, in step 3, by applying a hierarchical clustering algorithm, each edge server creates a corresponding identity template for each terminal based on the received information. The template consists of a unique identity identifier and corresponding identity credentials, and uses the similarity of behavior patterns to make a comprehensive decision. a. Edge topology map construction: For terminals with similar patterns, create corresponding identity maps for them and build a topology map based on the results. ; The node set is recorded as ,in Represents the terminals in the cluster, } ,in Represents an edge between two terminals; The scale of identity credential representation generation for all terminals in the same decision cluster is The identity metric matrix of in, represents the dimension of the identity credential vector, Represents the number of registered terminals in the same cluster; Connection Credentials Vector and The edge between Represents an edge The weight (distance) is: (1) in and Represents vectors and of dimensional identity characteristics; The edge server sorts the weights within the cluster from high to low, expressed as , here Represents the number of terminals in the same decision cluster, is the terminal weight that is closest to a specific authentication terminal mode, is the terminal weight that deviates most from a specific authenticated terminal pattern; these credential vectors are stored in the edge server for further authentication; b. Collaborative verification between individuals and collective intelligence: In the verification phase, A terminal generates new behavior data. For the privacy-sensitive data of the terminal, the input data is trained and generated by the feature extractor, and the generated vector representation is recorded as , and sent to the nearest edge server for verification; When the corresponding edge server receives the corresponding authentication request, it will make a two-stage authentication decision based on the collaborative authentication committee mechanism. That is, it uses the similarity of behavior patterns to make a comprehensive decision. Specifically, it includes two stages: individual authentication decision and group authentication decision: Individual authentication decision: In the first stage authentication process, the input vector The behavior template stored in the terminal The distance between the input behavior data and the behavior template is measured by the change in the intensity relationship; According to the distance formula, we can get the edge Weight , for The percentage change in intensity relationship is calculated using the following formula: : in Represents the weight of one's own identity vector, if A positive value indicates that the behaviors are closer, otherwise it indicates that they are more divergent; Decision of crowd authentication: In the second phase of verification, the input vector is measured The change in the strength relationship compared with other terminals in the original topology. Specifically, the generated behavior representation vector is transmitted to the edge server, and a new matrix copy is created to replace the original identity embedding vector. The existing identity credential vector is recalculated based on the previous distance formula with other terminals in the same decision cluster. The change in the intensity relationship between ; The edge server stores the input vector and the weights of other users in the same decision cluster, denoted as ; For users to be authenticated , firstly, the connection strength is adjusted by SoftMax Normalize and get the corresponding weights , and thus calculate the change in the percentage of connection strength within the same decision cluster according to the following formula: like The increase in intensity indicates that the new behavior vector tends to shorten the distance between the behavior pattern to be authenticated and the users with similar behavior patterns in the same decision cluster; otherwise, if A decrease indicates that the new behavior vector tends to shift the distance between the behavior pattern to be authenticated and users with similar behavior patterns in the same decision cluster. This also indicates that the behavior pattern is likely initiated by an attacker. c. Adaptive decision aggregation: Design an adaptive two-stage verification weight aggregation method; for the The terminal to be authenticated, if the distribution of its characterization vector is recorded as , the corresponding variance is recorded as ; The distribution within a decision cluster is recorded as , the corresponding variance is recorded as ;definition The adaptive parameters are To weight the two-stage authentication results, the calculation formula is as follows: therefore The weight of Therefore, the final authentication result combines the two-stage decision-making. In order to further adaptively integrate the results of individual decision-making and group decision-making, and considering the gap between the distributions, the final authentication result is fused according to the following formula: The process is formally described as follows: , we can get ;otherwise ;in Represents the final authentication result returned by the edge-deployed authentication server. Represents the authentication threshold.

Images