Power monitoring system ai model anomaly detection method and system based on trusted computing

Abstract

The application discloses a kind of based on trusted computing's electric power monitoring system AI model anomaly detection method and system, method includes: in electric power monitoring system, construct trusted execution environment, securely load and run AI model;Model loading, input and output are whole chain credibility measurement and check;Real-time acquisition electric power data, model internal feature and system resource feature, carry out multidimensional fusion and standardization;Using the model based on contrast learning training, fusion feature is mapped to embedding space, calculate its similarity with normal state reference vector, determine operation anomaly accordingly;Based on integrity hash and output credibility, generate comprehensive credibility index, trigger hierarchical alarm and protection;While in trusted environment, using abnormal sample to carry out safe incremental update to detection model.The application realizes the whole life cycle trusted monitoring and active safety protection to AI model operating state, significantly improves the ability of electric power monitoring system to resist tampering and unknown attack.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for power systems, and particularly relates to an anomaly detection method and system for AI models in power monitoring systems based on trusted computing. Background Technology

[0002] With the rapid advancement of intelligence and digitalization in power systems, artificial intelligence models are widely used in critical tasks such as load forecasting, fault diagnosis, and safe dispatching in power monitoring. These models significantly improve the automation level and decision-making efficiency of power grid operation by analyzing massive amounts of real-time data. However, the inherent "black box" nature of AI models leads to a lack of transparency and interpretability in their decision-making process, introducing uncertainty and security risks in practical applications. For example, models may lack robustness to abnormal input data or exhibit severe deviations when subjected to adversarial attacks or parameter tampering, thereby jeopardizing the safe and stable operation of the power system.

[0003] Currently, the protection of AI models in power monitoring systems mainly relies on external measures, such as input data verification, output result rationality analysis, or the deployment of traditional intrusion detection systems. While these methods can detect some explicit errors, they cannot reliably monitor the internal operating state of the model and are insufficient to reveal hidden anomalies in the reasoning process. Once the model is maliciously tampered with or attacked, existing methods often only respond passively after the damage occurs, lacking the ability to prevent problems before they occur or to block them during the attack. Furthermore, the systems generally lack traceable and verifiable security mechanisms based on trusted computing, making it difficult to trace anomalies and further increasing operational risks.

[0004] Therefore, there is an urgent need for a technical solution that can deeply ensure the safety of AI model operation, realize reliable monitoring throughout the process, and handle anomalies in a timely manner. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and provide a method and system for anomaly detection of AI models in power monitoring systems based on trusted computing. This method constructs a trusted execution environment to implement integrity protection and behavior monitoring of the AI ​​model throughout its entire lifecycle. Combined with multi-dimensional feature fusion and intelligent anomaly detection, it achieves real-time perception of the model's operating status, accurate anomaly judgment, and adaptive security enhancement, thereby comprehensively improving the intelligent security defense level of the power monitoring system.

[0006] In a first aspect, the present invention provides an anomaly detection method for an AI model in a power monitoring system based on trusted computing, comprising: A trusted execution environment is constructed in the power monitoring system, and the AI ​​model is loaded into the trusted execution environment to run. During the operation of the AI ​​model, the weight parameters, input data and inference output of the AI ​​model are fully quantified and verified for completeness, a measurement result is generated, and it is determined whether the measurement result is consistent with the preset benchmark value. If consistent, real-time collection of power operation data, internal features of the AI ​​model, and system resource characteristics is performed, and these are standardized and integrated to form a feature vector reflecting the operating status of the system and the AI ​​model. ; The feature vector Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If so, the AI ​​model is determined to be in an abnormal operating state; After detecting an abnormal running state of the AI ​​model, based on the input data integrity hash... AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

[0007] Secondly, the present invention provides an anomaly detection system for an AI model in a power monitoring system based on trusted computing, comprising: The module is configured to build a trusted execution environment in the power monitoring system and load the AI ​​model into the trusted execution environment for execution. The generation module is configured to perform full-scale measurement and integrity verification on the weight parameters, input data and inference output of the AI ​​model during the operation of the AI ​​model, generate measurement results, and determine whether the measurement results are consistent with the preset benchmark value. The fusion module, configured to collect real-time power operation data, internal features of the AI ​​model, and system resource features if they are consistent, then standardizes and merges them to form a feature vector reflecting the operating status of the system and the AI ​​model. ; The calculation module is configured to process the feature vector. Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If so, the AI ​​model is determined to be in an abnormal operating state; The early warning module is configured to, upon detecting an abnormal running state of the AI ​​model, trigger an alert based on the integrity hash of the input data. AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

[0008] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the AI ​​model anomaly detection method for power monitoring system based on trusted computing according to any embodiment of the present invention.

[0009] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the steps of the AI ​​model anomaly detection method for power monitoring system based on trusted computing according to any embodiment of the present invention.

[0010] The anomaly detection method and system for AI models in power monitoring systems based on trusted computing disclosed in this application have the following specific advantages: Trusted computing technology has built a hardware-level secure isolation environment for AI models, which fundamentally prevents model parameters and running data from being tampered with or stolen, ensuring the trustworthy foundation of the running process. By deeply integrating power operation data, model internal state, and system resource characteristics, and using a comparative learning mechanism to construct a normal behavior benchmark, accurate and early detection of latent anomalies in the model is achieved. A quantitative evaluation index covering integrity and behavioral credibility was designed, and the detection and adaptive update mechanism was deeply integrated into the trusted environment, forming a proactive security closed loop of "monitoring-evaluation-protection-evolution", which significantly improves the system's continuous defense capability against unknown threats. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart of an anomaly detection method for an AI model in a power monitoring system based on trusted computing, provided as an embodiment of the present invention; Figure 2 This is a structural block diagram of an AI model anomaly detection system for power monitoring based on trusted computing, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Please see Figure 1 The diagram illustrates a flowchart of an anomaly detection method for an AI model in a power monitoring system based on trusted computing, as described in this application.

[0015] like Figure 1 As shown, the anomaly detection method for AI models in power monitoring systems based on trusted computing specifically includes the following steps: Step S101: Construct a trusted execution environment in the power monitoring system and load the AI ​​model into the trusted execution environment for operation.

[0016] In this step, during device startup, a boot metric chain is formed by measuring the BIOS, bootloader, operating system kernel, and monitoring applications level by level through a root of trust. And write the metric value to the platform configuration register, where, This is a binary image file for the BIOS. This is a binary image file of the operating system. For the application's binary image file, A cryptographic hash function, used to map input data of arbitrary length to a fixed-length 256-bit feature value; When the AI ​​model is loaded into the trusted execution environment, the weight parameters of the AI ​​model are calculated. hash value And compare it with the pre-stored baseline signature value; The measurement results of the platform configuration register are signed using a remote proof mechanism. Send to a remote server for verification. The platform configuration registers are immutable registers in trusted computing hardware used to store system metrics. This refers to a private key signature algorithm, which means using an asymmetric encryption private key generated within a trusted execution environment to digitally sign data. Digital signature results generated for a trusted execution environment.

[0017] Step S102: During the operation of the AI ​​model, the weight parameters, input data and inference output of the AI ​​model are fully quantified and verified for completeness, a measurement result is generated, and it is determined whether the measurement result is consistent with the preset benchmark value.

[0018] In this step, if the measurement result is inconsistent with the preset benchmark value, the AI ​​model will be prevented from running or a first alarm will be triggered.

[0019] Step S103: If consistent, real-time collection of power operation data, internal features of the AI ​​model, and system resource features is performed, and these are standardized and integrated to form a feature vector reflecting the operating status of the system and the AI ​​model. .

[0020] In this step, the current is collected. ,Voltage ,frequency Power factor Constructing power operation data vectors ; Extracting latent vectors from the intermediate layers of the AI ​​model ; Collect CPU utilization Memory consumption Disk I / O frequency ; Vector of power operation data Latent vectors CPU utilization Memory consumption and disk I / O frequency Merge into a unified feature vector and the unified feature vector The feature vector is obtained by standardization. The expression is: , In the formula, Let be the normalized value of the i-th feature in the fused feature vector at time t. Let i be the i-th dimension component of the original fused feature vector acquired at time t. Let be the mean of the i-th feature. Let be the standard deviation of the i-th feature.

[0021] Step S104, the feature vector Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If so, the AI ​​model is determined to be in an abnormal operating state.

[0022] In this step, deep neural networks are used. The normalized feature vectors are mapped to the embedding space to obtain low-dimensional embedding vectors. The expression is , wherein, used for training the deep neural network loss function for: , In the formula, For the set of positive sample pairs, For the set of negative sample pairs, This is the interval threshold. This represents the similarity between positive sample pairs in the embedding space. The embedding similarity of negative sample pairs in the embedding space; The reference embedding vector is calculated based on historical normal data, and its expression is: , In the formula, As a reference embedding vector, This represents the normal sample size. Compute low-dimensional embedding vectors Reference embedding vector in normal state similarity .

[0023] Step S105: After detecting an abnormal running state of the AI ​​model, based on the input data integrity hash... AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

[0024] In this step, a comprehensive credibility index is used. The calculation formula is: , In the formula, , , Input data integrity hash Weighting coefficients, AI model parameter integrity hashing Weighting coefficients and output credibility hash The weighting coefficients.

[0025] Set a minimum confidence threshold .like This indicates a severe lack of system credibility, immediately triggering a high-level alarm. Simultaneously, the system automatically performs protective actions, such as isolating abnormal outputs from the current model, blocking suspicious data streams, or switching to a pre-prepared secure backup model, ensuring uninterrupted power monitoring operations.

[0026] Specifically, implementing protective measures includes adaptive updates, specifically: Anomalies are stored in a secure storage area managed by a trusted execution environment to form an anomaly dataset. ; An incremental learning method is used, combined with historical normal datasets. With the aforementioned abnormal dataset Construct a new training set And by optimizing the loss function Update the model parameters, where, This is the balance coefficient; The parameter update is performed in a trusted execution environment to obtain the new model parameters. ,in, The learning rate; For new model parameters Hash calculations are performed and the results are sent to the verification server via a remote verification mechanism. Once verified, the results are updated synchronously across all network monitoring nodes.

[0027] In summary, the method of this application, by introducing a hardware-level trusted execution environment, constructs a tamper-proof and verifiable security foundation for the entire operation of the AI ​​model, fundamentally eliminating the risk of malicious damage or theft of parameters and data during loading and inference. Furthermore, by innovatively integrating multi-dimensional information such as real-time power operation data, deep features within the AI ​​model, and system resource status, and utilizing a comparative learning mechanism to construct a feature space and dynamic reference benchmark that accurately characterizes normal behavior patterns, it achieves early and accurate identification of hidden operational anomalies caused by adversarial attacks, data poisoning, or self-faults, significantly improving the comprehensiveness and sensitivity of anomaly detection. Based on this... By designing a quantitative evaluation index that integrates integrity verification and behavioral credibility, and forming an organic closed loop of anomaly detection, security assessment, real-time protection, and incremental learning updates completed in a trusted environment, this system can not only quickly trigger proactive protection strategies such as alarms, isolation, or switching when threats are detected to ensure business continuity, but also enable the detection model to continuously evolve using newly discovered anomaly samples, thereby dynamically improving its adaptive defense capabilities against unknown attack patterns and new anomalies. Ultimately, this system constructs a full lifecycle proactive security defense system covering pre-event prevention, in-event monitoring, post-event handling, and continuous evolution, greatly enhancing the inherent security and operational reliability of the power monitoring system in the process of intelligentization.

[0028] Please see Figure 2 The diagram shows a structural block diagram of an AI model anomaly detection system for power monitoring based on trusted computing, as described in this application.

[0029] like Figure 2 As shown, the AI ​​model anomaly detection system 200 for power monitoring systems includes a construction module 210, a generation module 220, a fusion module 230, a calculation module 240, and an early warning module 250.

[0030] The system includes a construction module 210 configured to build a trusted execution environment in the power monitoring system and load the AI ​​model into the trusted execution environment for operation; a generation module 220 configured to perform full-scale measurement and integrity verification on the weight parameters, input data, and inference output of the AI ​​model during its operation, generate measurement results, and determine whether the measurement results are consistent with preset benchmark values; and a fusion module 230 configured to, if consistent, collect power operation data, internal features of the AI ​​model, and system resource features in real time, and perform standardization and fusion to form a feature vector reflecting the operating status of the system and the AI ​​model. ; Calculation module 240, configured to calculate the feature vector Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If the AI ​​model's operating state is found to be abnormal, the early warning module 250 is configured to, upon detecting an abnormal AI model operating state, determine an abnormality based on the input data integrity hash. AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

[0031] It should be understood that Figure 2 The modules and references described in the document Figure 1 The steps described in the text correspond to those in the method described above. Therefore, the operations, features, and corresponding technical effects described above also apply to the method described in the text. Figure 2 The various modules in the document will not be described in detail here.

[0032] In other embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor performs the AI ​​model anomaly detection method for power monitoring system based on trusted computing in any of the above method embodiments. In one embodiment, the computer-readable storage medium of the present invention stores computer-executable instructions, which are configured as follows: A trusted execution environment is constructed in the power monitoring system, and the AI ​​model is loaded into the trusted execution environment to run. During the operation of the AI ​​model, the weight parameters, input data and inference output of the AI ​​model are fully quantified and verified for completeness, a measurement result is generated, and it is determined whether the measurement result is consistent with the preset benchmark value. If consistent, real-time collection of power operation data, internal features of the AI ​​model, and system resource characteristics is performed, and these are standardized and integrated to form a feature vector reflecting the operating status of the system and the AI ​​model. ; The feature vector Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If so, the AI ​​model is determined to be in an abnormal operating state; After detecting an abnormal running state of the AI ​​model, based on the input data integrity hash... AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

[0033] Computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the AI ​​model anomaly detection system for a trusted computing-based power monitoring system. Furthermore, the computer-readable storage medium may include high-speed random access memory, and may also include memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include memory remotely configured relative to a processor, which can be connected to the AI ​​model anomaly detection system for a trusted computing-based power monitoring system via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0034] Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiment of the present invention, such as... Figure 3 As shown, the device includes a processor 310 and a memory 320. The electronic device may also include an input device 330 and an output device 340. The processor 310, memory 320, input device 330, and output device 340 can be connected via a bus or other means. Figure 3 Taking a bus connection as an example, the memory 320 is the computer-readable storage medium described above. The processor 310 executes various server functions and data processing by running non-volatile software programs, instructions, and modules stored in the memory 320, thereby implementing the above-described method embodiment of the AI ​​model anomaly detection method for a power monitoring system based on trusted computing. The input device 330 can receive input digital or character information and generate key signal inputs related to user settings and function control of the AI ​​model anomaly detection system for a power monitoring system based on trusted computing. The output device 340 may include a display screen or other display device.

[0035] The aforementioned electronic device can execute the method provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in the embodiments of the present invention.

[0036] In one implementation, the aforementioned electronic device is applied to an AI model anomaly detection system for a power monitoring system based on trusted computing, serving as a client, and includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: A trusted execution environment is constructed in the power monitoring system, and the AI ​​model is loaded into the trusted execution environment to run. During the operation of the AI ​​model, the weight parameters, input data and inference output of the AI ​​model are fully quantified and verified for completeness, a measurement result is generated, and it is determined whether the measurement result is consistent with the preset benchmark value. If consistent, real-time collection of power operation data, internal features of the AI ​​model, and system resource characteristics is performed, and these are standardized and integrated to form a feature vector reflecting the operating status of the system and the AI ​​model. ; The feature vector Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If so, the AI ​​model is determined to be in an abnormal operating state; After detecting an abnormal running state of the AI ​​model, based on the input data integrity hash... AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

[0037] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0038] 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 them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An anomaly detection method for an AI model in a power monitoring system based on trusted computing, characterized in that, include: A trusted execution environment is constructed in the power monitoring system, and the AI ​​model is loaded into the trusted execution environment to run. During the operation of the AI ​​model, the weight parameters, input data and inference output of the AI ​​model are fully quantified and verified for completeness, a measurement result is generated, and it is determined whether the measurement result is consistent with the preset benchmark value. If consistent, real-time collection of power operation data, internal features of the AI ​​model, and system resource characteristics is performed, and these are standardized and integrated to form a feature vector reflecting the operating status of the system and the AI ​​model. Specifically, it includes: Current sampling ,Voltage ,frequency Power factor Constructing power operation data vectors ; Extracting latent vectors from the intermediate layers of the AI ​​model ; Collect CPU utilization Memory consumption Disk I / O frequency ; Vector of power operation data Latent vectors CPU utilization Memory consumption and disk I / O frequency Merge into a unified feature vector and the unified feature vector The feature vector is obtained by standardization. The expression is: ， In the formula, Let be the normalized value of the i-th feature in the fused feature vector at time t. Let i be the i-th dimension component of the original fused feature vector acquired at time t. Let be the mean of the i-th feature. Let be the standard deviation of the i-th feature; The feature vector Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If the AI ​​model is found to be in an abnormal operating state, then the low-dimensional embedding vector is calculated. Reference embedding vector in normal state similarity include: Through deep neural networks The normalized feature vectors are mapped to the embedding space to obtain low-dimensional embedding vectors. The expression is , wherein, used for training the deep neural network loss function for: ， In the formula, For the set of positive sample pairs, For the set of negative sample pairs, This is the interval threshold. This represents the similarity between positive sample pairs in the embedding space. The embedding similarity of negative sample pairs in the embedding space; The reference embedding vector is calculated based on historical normal data, and its expression is: ， In the formula, As a reference embedding vector, This represents the normal sample size. Compute low-dimensional embedding vectors Reference embedding vector in normal state similarity ; After detecting an abnormal running state of the AI ​​model, based on the input data integrity hash... AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

2. The anomaly detection method for an AI model in a power monitoring system based on trusted computing according to claim 1, characterized in that, The construction of a trusted execution environment in the power monitoring system includes: During device startup, a boot metric chain is formed by measuring the BIOS, bootloader, operating system kernel, and monitoring applications step by step through a root of trust. And write the metric value to the platform configuration register, where, This is a binary image file for the BIOS. This is a binary image file of the operating system. For the application's binary image file, A cryptographic hash function, used to map input data of arbitrary length to a fixed-length 256-bit feature value; When the AI ​​model is loaded into the trusted execution environment, the weight parameters of the AI ​​model are calculated. hash value And compare it with the pre-stored baseline signature value; The measurement results of the platform configuration register are signed using a remote proof mechanism. Send to a remote server for verification. The platform configuration registers are immutable registers in trusted computing hardware used to store system metrics. This refers to a private key signature algorithm, which means using an asymmetric encryption private key generated within a trusted execution environment to digitally sign data. Digital signature results generated for a trusted execution environment.

3. The anomaly detection method for an AI model in a power monitoring system based on trusted computing according to claim 1, characterized in that, The comprehensive credibility index The calculation formula is: ， In the formula, , , Input data integrity hash Weighting coefficients, AI model parameter integrity hashing Weighting coefficients and output credibility hash The weighting coefficients.

4. The anomaly detection method for an AI model in a power monitoring system based on trusted computing according to claim 1, characterized in that, in, Implementing protective measures includes adaptive updates, specifically: Anomalies are stored in a secure storage area managed by a trusted execution environment to form an anomaly dataset. ; An incremental learning method is used, combined with historical normal datasets. With the aforementioned abnormal dataset Construct a new training set And by optimizing the loss function Update the model parameters, where, This is the balance coefficient; The parameter update is performed in a trusted execution environment to obtain the new model parameters. ,in, The learning rate; For new model parameters Hash calculations are performed and the results are sent to the verification server via a remote verification mechanism. Once verified, the results are updated synchronously across all network monitoring nodes.

5. An anomaly detection system for an AI model in a power monitoring system based on trusted computing, characterized in that, include: The module is configured to build a trusted execution environment in the power monitoring system and load the AI ​​model into the trusted execution environment for execution. The generation module is configured to perform full-scale measurement and integrity verification on the weight parameters, input data and inference output of the AI ​​model during the operation of the AI ​​model, generate measurement results, and determine whether the measurement results are consistent with the preset benchmark value. The fusion module, configured to collect real-time power operation data, internal features of the AI ​​model, and system resource features if they are consistent, then standardizes and merges them to form a feature vector reflecting the operating status of the system and the AI ​​model. Specifically, it includes: Current sampling ,Voltage ,frequency Power factor Constructing power operation data vectors ; Extracting latent vectors from the intermediate layers of the AI ​​model ; Collect CPU utilization Memory consumption Disk I / O frequency ; Vector of power operation data Latent vectors CPU utilization Memory consumption and disk I / O frequency Merge into a unified feature vector and the unified feature vector The feature vector is obtained by standardization. The expression is: ， In the formula, Let be the normalized value of the i-th feature in the fused feature vector at time t. Let i be the i-th dimension component of the original fused feature vector acquired at time t. Let be the mean of the i-th feature. Let be the standard deviation of the i-th feature; The calculation module is configured to process the feature vector. Mapped to low-dimensional embedding vectors And compute the low-dimensional embedding vector. Reference embedding vector in normal state similarity If similarity Below the set first threshold If the AI ​​model is found to be in an abnormal operating state, then the low-dimensional embedding vector is calculated. Reference embedding vector in normal state similarity include: Through deep neural networks The normalized feature vectors are mapped to the embedding space to obtain low-dimensional embedding vectors. The expression is , wherein, used for training the deep neural network loss function for: ， In the formula, For the set of positive sample pairs, For the set of negative sample pairs, This is the interval threshold. This represents the similarity between positive sample pairs in the embedding space. The embedding similarity of negative sample pairs in the embedding space; The reference embedding vector is calculated based on historical normal data, and its expression is: ， In the formula, As a reference embedding vector, This represents the normal sample size. Compute low-dimensional embedding vectors Reference embedding vector in normal state similarity ; The early warning module is configured to, upon detecting an abnormal running state of the AI ​​model, trigger an alert based on the integrity hash of the input data. AI model parameter integrity hashing and output a trustworthy hash Generate a comprehensive credibility index If we consider the overall credibility index Below the preset second threshold If this occurs, a second alarm will be triggered and protective measures will be implemented.

6. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 4.

Images