Intelligent power plant trust evaluation method based on deep learning
By adopting a deep learning-based trust evaluation method in smart power plants, using Transformer model and particle swarm optimization algorithm for trust evaluation, the problem that traditional network security measures cannot adapt to dynamic changes is solved, and a more accurate and dynamic trust evaluation is achieved.
Patent Information
- Application Number
- CN202411695907.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-25
- Publication Date
- 2025-02-14
AI Technical Summary
Traditional cybersecurity measures cannot adapt to the dynamic changes and complex needs of smart power plant networks, resulting in attackers being able to control the entire network through horizontal scaling.
The smart power plant trust evaluation method based on deep learning is adopted, and the Transformer model and particle swarm optimization algorithm of the selective attention mechanism are used to perform selective ensemble learning to predict the trusted score of the access subject.
It realizes dynamic trust assessment of network requests and activities, can reflect changes in subject behavior and adapt to unknown threats, reduces the risk of misjudgment and improves the accuracy of trust assessment.
Smart Images
Figure CN119443970A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a deep learning-based trust assessment method for a smart power plant, and belongs to the technical field of network security. Background Art
[0002] With the rapid development of new generation technologies, security risks and attack incidents continue to emerge under the new technological situation. Traditional network security measures mainly rely on preset rules and static protection methods, which can no longer adapt to the dynamic changes and complex needs of smart power plant networks. For example, once an attacker obtains access rights to a host in the intranet, and because the existing defense strategy has relatively loose control over intranet access devices, the attacker can expand horizontally in the intranet through a series of operations and eventually control the entire network. In this case, even if the network adopts perfect rules and protection methods, it cannot resist attacks. In the final analysis, it is because the security system trusts the devices that have been connected to the network too much and fails to achieve dynamic protection. Under the continuous dynamic trust assessment management system based on deep learning, trust scores are generated by continuously conducting trust assessments on various entities connected to the network, and different control measures are taken according to the threshold of the trust scores obtained.
[0003] In this management system, the trust assessment engine is the core component, which is used to numerically evaluate the credibility of requests and activities issued by various access entities in the network. The access control engine makes further authorization decisions based on this trust assessment to determine whether the current device has security risks and what management measures should be taken. How to conduct reasonable trust assessments on network requests and activities is the first problem to be solved when the management system is implemented.
[0004] In the existing trust assessment literature, for example, the traditional trust assessment method based on rules and policies relies on predefined rules and policies. Although it is simple to implement, as a static assessment method, it cannot timely reflect changes in subject behavior and adapt to unknown threats. In addition, it is difficult to define rules and is prone to missed judgments and misjudgments.
[0005] Therefore, the present invention proposes a trust assessment method for a smart power plant based on deep learning. The method adopts selective ensemble learning, in which the Transformer model with selective attention mechanism is used as the basic classifier, and the particle swarm optimization algorithm (PSO) is used to obtain the optimal aggregation weight vector to predict the trust score of the access subject. This algorithm has higher robustness. Summary of the invention
[0006] The present invention provides a smart power plant trust assessment method based on deep learning. It uses trust features to predict the trust score of the access subject. The method can be used in the smart power plant network to continuously and dynamically perform trust assessment on each subject accessing the network. The obtained trust score will be used for subsequent access control.
[0007] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows:
[0008] A smart power plant trust assessment method based on deep learning, including:
[0009] In the first stage, the access subject is the user-device pair, which obtains the data information of the user and the device from the trusted environment perception system and preprocesses the data information. The characteristic value is represented by q ij , represents the jth feature value of the i-th request, which is stored in the data storage system; then the direct evaluation model uses the feature value in the data storage system as input to calculate the direct trust score; then, based on the device association relationship, the indirect evaluation model is used to calculate the access behavior of the access subject at this time from the perspective of adjacent associated devices using the collected user information and part of the device information to obtain the indirect trust score; both the direct evaluation model and the indirect evaluation model are the Transformer models described below, but the training data are different;
[0010] The structure of the Transformer model used in data preprocessing, direct evaluation model, and indirect evaluation model:
[0011] For the training data set D = {(q i1 ,q i2 ,...,q ij ),T|F}, where T and F are manually labeled as credible and uncredible respectively, q ij is the feature value, indicating the jth feature value of the i-th request; it is randomly divided into a training set and a test set in a ratio of 7:3, which are used to train the Transformer model and evaluate the model performance respectively;
[0012] Transformer input layer design: User data embedding: embed all categorical features of users, and combine these embedded vectors and numerical features into a complete user feature vector; Device data embedding: embed the categorical features of devices, and finally combine these embedded vectors and numerical features into a complete device feature vector; then merge the user feature vector and the device feature vector through the fusion layer to form a complete input vector;
[0013] Among them, the selective attention mechanism uses the Transformer's multi-head self-attention mechanism (Multi-Head Attention) to process the fused feature vector, thereby capturing the relationship between user and device features. Through the selective attention mechanism, we can focus on the features that are most important for trust assessment;
[0014] In the second stage, the adaptive weights are optimized using the search of the particle swarm optimization algorithm (PSO);
[0015] In the third stage, the adaptive weights optimized in the second stage are used, combined with the direct trust score, indirect trust score, and historical trust score predicted in the first stage, to calculate the final trust score of the access subject, and simultaneously update the historical trust score of the node.
[0016] In the first stage above, the access subject is not an isolated single user or a single device, but a user-device pair.
[0017] The above-mentioned user data information includes: user identification, user's current login location, user's historical login locations, user authentication information, user enhanced authentication information, number of user authentication failures, user activity, user active cycle, user request frequency and number of user unauthorized access; device data information includes: device identification, device type, antivirus protection component information, system high-risk vulnerability information and operating system version.
[0018] In the second stage, the data set of the particle swarm algorithm is Q = {Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in}, where Q id represents the direct trust score obtained from the i-th access request, Q ii represents the indirect trust score obtained from the i-th access request, Q in There are n in total, which represents the nearest n historical credibility scores participating in the i-th credibility score evaluation. The specific steps for optimizing the adaptive weights are as follows:
[0019] Step 201, define an objective function, which is used to evaluate the quality of weight distribution, wherein the objective function is as follows:
[0020] F(w)=f(Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in )
[0021] Where w is the weight vector, w = [w_Q id ,w_Q ii ,w_Qi1 ,w_Q i2 ,...,w_Q in ], then optimize F(w);
[0022] Step 202, initialize the particle swarm, each particle represents a possible weight vector w, the dimension of each particle is n+2, and initialize the velocity v and position x of each particle, and randomly assign a weight value;
[0023] Step 203, the speed is set according to v i (t+1)=m·v i (t)+c 1 ·r 1 ·(p i (t)-x i (t))+c 2 ·r 2 ·(g(t)-x i (t)) update, where m is the inertia weight, c 1 、c 2 is the learning factor, r 1 、r 2 is a random number, p i (t) is the current optimal position of the particle, g(t) is the global optimal position; the position is calculated according to x i (t+1)=x i (t)+v i (t+1) update, where x i (t) is the current weight vector of the particle;
[0024] Step 204, evaluate the particles, use the objective function F(w) to calculate the current fitness value of each particle, and update the local optimal position p of each particle i (t) and the global optimal position g(t);
[0025] Step 205, repeating the above steps 203 and 204 until the maximum number of iterations is reached or the convergence condition is met;
[0026] Step 206, after iteration, an optimal weight vector w is obtained, which combines the parameters
[0027] {Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in} is optimized in terms of weight value distribution.
[0028] In the third stage above, the optimal weight vector w obtained in the second stage is used o , combined with {Qid,Qii,Qi 1,Qi 2 ,...,Qin} According to the formula: Q im =w id ×Q id +w ii ×Q ii +w i1 ×Q i1 +w i2 ×Q i2 +...+w in ×Q in Obtain the final credibility score Q of the i-th request im , and update the historical credibility score.
[0029] The technologies not mentioned in the present invention are all referred to the prior art.
[0030] The trust assessment method for smart power plants based on deep learning in this invention adopts selective ensemble learning, uses the Transformer model based on the selective attention mechanism as the basic classifier, uses the particle swarm optimization algorithm to obtain the most aggregated weight vector, selects the relevant attributes of users and devices as trust features, and outputs the trust score based on the features. Compared with traditional trust assessment methods, it solves the problems of missed judgments and wrong judgments caused by the inability to reflect changes in subject behavior, the inability to adapt to unknown threats, and definition difficulties, and has better accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 This is a schematic diagram of the trust assessment method for a smart power plant according to the present invention;
[0032] Figure 2 The trust evaluation framework based on particle swarm optimization algorithm and Transformer model described in the present invention; DETAILED DESCRIPTION
[0033] In order to better understand the present invention, the content of the present invention is further explained below in conjunction with the embodiments, but the content of the present invention is not limited to the following embodiments.
[0034] This example provides a trust assessment method for a smart power plant based on deep learning. This method uses selective ensemble learning, uses the Transformer model based on the selective attention mechanism as the basic classifier, uses the particle swarm optimization algorithm to obtain the most aggregated weight vector, and selects the relevant attributes of users and devices as trust features.
[0035] Figure 1 This is a schematic diagram of the smart power plant trust assessment method described in the present invention.
[0036] The access subject is not an isolated single user or single device, but a user-device pair. When conducting trust assessment on the access subject, first of all, it is necessary to obtain relevant data of the user and the device from the trusted environment perception system. User data information includes: user identification, user current login location, user historical login location, user authentication information, user enhanced authentication information, number of user authentication failures, user activity, user active cycle, user request frequency, number of user unauthorized access; device data information includes: device identification, device type, antivirus protection component information, system high-risk vulnerability information, operating system version; secondly, the data is processed, and the characteristic value is expressed as q ij , represents the jth feature value of the i-th request, which is stored in the data storage system; then the direct evaluation model uses the feature value in the data storage system as input to calculate the direct trust score; finally, based on the device association relationship, the indirect evaluation model is used to calculate the access behavior of the access subject at this time from the perspective of adjacent associated devices using the collected user information and part of the device information to obtain the indirect trust score.
[0037] Figure 2 This is the trust assessment framework based on the particle swarm optimization algorithm and the Transformer model described in the present invention.
[0038] The first stage includes data preprocessing, direct evaluation model and structural design of the Transformer model used in the indirect evaluation model, which is used to predict the direct credibility score and indirect credibility score of the access subject respectively;
[0039] In the second stage, the adaptive weights are optimized using the search of the particle swarm optimization algorithm (PSO);
[0040] In the third stage, the adaptive weights optimized in the second stage are used, combined with the direct trust score, indirect trust score, and historical trust score predicted in the first stage, to calculate the final trust score of the access subject, and simultaneously update the historical trust score of the node.
[0041] In the first stage mentioned above, the network structure and input and output formats of the Transformer model are designed according to the requirements of the trust score prediction problem. Among them, the access subject is not an isolated single user or a single device, but a user-device pair. When conducting trust evaluation on the access subject, it is first necessary to obtain relevant data of the user and the device from the trusted environment perception system. The user data information includes: user identification, user current login location, user historical login location, user authentication information, user enhanced authentication information, number of user authentication failures, user activity, user active period, user request frequency, number of user unauthorized access; device data information includes: device identification, device type, antivirus protection component information, system high-risk vulnerability information, operating system version; secondly, the data is processed, and the characteristic value is represented by qij , represents the jth feature value of the i-th request, which is stored in the data storage system; then the direct evaluation model uses the feature value in the data storage system as input to calculate the direct trust score; finally, based on the device association relationship, the indirect evaluation model is used to calculate the access behavior of the access subject at this time from the perspective of adjacent associated devices using the collected user information and part of the device information to obtain the indirect trust score.
[0042] For the training data set D = {(q i1 ,q i2 ,...,q ij ),T|F}, where T and F are manually labeled as credible and uncredible respectively; they are randomly divided into training set and test set in a ratio of 7:3, which are used to train the Transformer model and evaluate the model performance respectively.
[0043] Transformer input layer design. User data embedding: embed all categorical features of the user, and combine these embedded vectors and numerical features into a complete user feature vector; device data embedding: embed the categorical features of the device, and finally combine these embedded vectors and numerical features into a complete device feature vector; then merge the user feature vector and the device feature vector through the fusion layer to form a complete input vector.
[0044] Among them, the selective attention mechanism uses the Transformer's multi-head self-attention mechanism (Multi-Head Attention) to process the fused feature vector, thereby capturing the relationship between user and device features. Through the selective attention mechanism, we can focus on the features that are most important for trust assessment.
[0045] In the second stage, the data set of the particle swarm algorithm is Q = {Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in}, where Q id represents the direct trust score obtained from the i-th access request, Q ii represents the indirect trust score obtained from the i-th access request, Q ij There are n in total, which represent the nearest n historical credibility scores participating in the i-th credibility score evaluation. The specific steps are as follows:
[0046] Step 201, define an objective function, which is used to evaluate the quality of weight distribution. The objective function is as follows:
[0047] F(w)=f(Qid ,Q ii ,Q i1 ,Q i2 ,...,Q in )
[0048] Where w is the weight vector, w = [w_Q id ,w_Q ii ,w_Q i1 ,w_Q i2 ,...,w_Q in ], then optimize F(w);
[0049] Step 202, initialize the particle swarm, each particle represents a possible weight vector w, the dimension of each particle is n+2, and initialize the velocity v and position x of each particle, and randomly assign a weight value;
[0050] Step 203, the speed is set according to v i (t+1)=m·v i (t)+c 1 ·r 1 ·(p i (t)-x i (t))+c 2 ·r 2 ·(g(t)-x i (t)) update, where m is the inertia weight, c 1 、c 2 is the learning factor, r 1 、r 2 is a random number, p i (t) is the current optimal position of the particle, g(t) is the global optimal position; the position is calculated according to x i (t+1)=x i (t)+v i (t+1) update, where x i (t) is the current weight vector of the particle;
[0051] Step 204: Evaluate the particles. Use the objective function F(w) to calculate the current fitness value of each particle and update the local optimal position p of each particle. i (t) and the global optimal position g(t);
[0052] Step 205, repeating the above steps 203 and 204 until the maximum number of iterations is reached or the convergence condition is met;
[0053] Step 206, after iteration, an optimal weight vector w is obtained, which combines the parameters
[0054] {Q id ,Qii ,Q i1 ,Q i2 ,...,Q in} is optimized in terms of weight value distribution.
[0055] In the third stage above, the optimal weight vector w obtained in the second stage is used o , combined with {Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in According to the formula: Q im =w id ×Q id +w ii ×Q ii +w i1 ×Q i1 +w i2 ×Q i2 +...+w in ×Q in Obtain the final credibility score Q of the i-th request im , and update the historical credibility score.
[0056] The above-mentioned smart power plant trust assessment method based on deep learning adopts selective ensemble learning, uses the Transformer model based on the selective attention mechanism as the basic classifier, adopts the particle swarm optimization algorithm to obtain the most aggregated weight vector, selects the relevant attributes of users and devices as trust features, and outputs the trust score based on the features. It can reflect the changes in subject behavior, adapt to unknown threats, and has no problems of missed or wrong judgments, with high accuracy.
[0057] The present invention provides a trust assessment method for a smart power plant based on deep learning. There are many methods and ways to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention. All components not specified in this embodiment can be implemented using existing technologies.
Claims
1. A smart power plant trust assessment method based on deep learning, characterized by: include: In the first stage, the access subject is the user-device pair, which obtains the data information of the user and the device from the trusted environment perception system and preprocesses the data information. The characteristic value is represented by q ij , represents the jth feature value of the i-th request, which is stored in the data storage system; The direct evaluation model then calculates the direct credibility score based on the feature values in the data storage system as input; Then, based on the device association relationship, using the collected user information and part of the device information, the indirect evaluation model is used to calculate the access behavior of the access subject at this time from the perspective of adjacent associated devices to obtain an indirect trust score; Among them, the design structure of the Transformer model used in data preprocessing, direct evaluation model and indirect evaluation model is as follows: For the training data set D = {(q i1 ,q i2 ,...,q ij ),T|F}, where T and F are manually labeled as credible and uncredible respectively, q ij is the feature value, indicating the jth feature value of the i-th request; it is randomly divided into a training set and a test set in a ratio of 7:3, which are used to train the Transformer model and evaluate the model performance respectively; Transformer input layer design: User data embedding: embed all categorical features of users, and combine these embedded vectors and numerical features into a complete user feature vector; Device data embedding: embed the categorical features of devices, and finally combine these embedded vectors and numerical features into a complete device feature vector; then merge the user feature vector and the device feature vector through the fusion layer to form a complete input vector; Among them, the selective attention mechanism uses the Transformer's multi-head self-attention mechanism to process the fused feature vector, thereby capturing the relationship between user and device features; In the second stage, the adaptive weights are optimized using the search of the particle swarm optimization algorithm; In the third stage, the adaptive weights optimized in the second stage are used, combined with the direct trust score, indirect trust score, and historical trust score predicted in the first stage, to calculate the final trust score of the access subject, and simultaneously update the historical trust score of the node.
2. The smart power plant trust assessment method based on deep learning according to claim 1, characterized in that: User data information includes: user identification, user's current login location, user's historical login locations, user authentication information, user enhanced authentication information, number of user authentication failures, user activity, user activity cycle, user request frequency and number of user unauthorized access; device data information includes: device identification, device type, antivirus protection component information, system high-risk vulnerability information and operating system version.
3. The smart power plant trust assessment method based on deep learning according to claim 1 or 2, characterized in that: In the second stage, the data set of the particle swarm algorithm is Q = {Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in }, where Q id represents the direct trust score obtained from the i-th access request, Q ii represents the indirect trust score obtained from the i-th access request, Q in There are n in total, which represents the nearest n historical credibility scores participating in the i-th credibility score evaluation. The specific steps for optimizing the adaptive weights are as follows: Step 201, define an objective function, which is used to evaluate the quality of weight distribution, wherein the objective function is as follows: F(w)=f(Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in ) Where w is the weight vector, w = [w_Q id ,w_Q ii ,w_Q i1 ,w_Q i2 ,...,w_Q in ], then optimize F(w); Step 202, initialize the particle swarm, each particle represents a possible weight vector w, the dimension of each particle is n+2, and initialize the velocity v and position x of each particle, and randomly assign a weight value; Step 203, the speed is set according to v i (t+1)=m·v i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(g(t)-x i (t)) update, where m is the inertia weight, c1, c2 are learning factors, r1, r2 are random numbers, p i (t) is the current optimal position of the particle, g(t) is the global optimal position; the position is calculated according to x i (t+1)=x i (t)+v i (t+1) update, where x i (t) is the current weight vector of the particle; Step 204, evaluate the particles, use the objective function F(w) to calculate the current fitness value of each particle, and update the local optimal position p of each particle i (t) and the global optimal position g(t); Step 205, repeating the above steps 203 and 204 until the maximum number of iterations is reached or the convergence condition is met; Step 206, after iteration, an optimal weight vector w is obtained, which combines the parameters {Q id ,Q ii ,Q i1 ,Q i2 ,...,Q in }'s weight value distribution is optimized.
4. The smart power plant trust assessment method based on deep learning according to claim 1 or 2, characterized in that: In the third stage, the optimal weight vector w obtained in the second stage is used o , combined with {Qid,Qii,Qi1,Qi2,...,Qin} according to the formula: Q im =w id ×Q id +w ii ×Q ii +w i1 ×Q i1 +w i2 ×Q i2 +...+w in ×Q in Get the final trust score Q of the i-th request im , and update the historical credibility score.