Analysis method for preventing AI data exception
By building an artificial intelligence anomaly feature graph model and using neural networks and spatiotemporal graph convolutional networks to determine data anomalies and conduct risk assessments, the difficult problem of artificial intelligence data anomaly management is solved, accurate early warning and dynamic defense are achieved, and the system's robustness and data security are improved.
Patent Information
- Application Number
- CN202510944580.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-09
- Publication Date
- 2025-09-19
AI Technical Summary
Existing technologies make it difficult to effectively prevent and manage AI data anomalies, leading to system disruptions and potential security risks.
By building an artificial intelligence abnormal feature map model, using autoencoders, RBM, DBN, DNN and other neural networks for data labeling and deep learning, combining spatiotemporal graph convolutional networks for abnormality judgment and early warning, and calculating risk levels based on ethical norms and legal rules for defense.
It achieves accurate identification and early warning of AI data anomalies, dynamically updates models to cope with abnormal evolution, and enhances system robustness and data security.
Smart Images
Figure CN120675784A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence, and in particular to an analysis method for preventing AI data anomalies. Background Art
[0002] Artificial intelligence (AI) is the study and development of technologies used to simulate humans, primarily manifesting in intelligent robotics and software. AI technology is now deeply embedded in our daily lives, particularly with the rapid development of large-scale models. It has moved from academic research to real-world applications, with widespread applications in healthcare, transportation, commerce, the military, and other fields, such as autonomous driving, facial recognition, intelligent recommendations, drones, and unmanned submarines. Data is ubiquitous, and AI relies on big data for data collection, mining, and analysis. In the realm of public data security, certain entities may infringe on personal privacy and corporate trade secrets. With the exponential advancement of information technology, AI will be able to automatically iterate and authorize itself. AI will be able to learn, think, judge, and update autonomously. AI systems will possess broad cognitive abilities comparable to or exceeding those of humans, capable of autonomous learning and reasoning across a wide range of complex tasks. This can lead to AI dysfunction, which can result in data manipulation and anomalies. Summary of the Invention
[0003] The purpose of the present invention is to provide an analysis method for preventing AI data anomalies to solve the problems raised in the prior art.
[0004] To achieve the above objectives, the present invention provides the following technical solutions: an analysis method for preventing AI data anomalies, and an intelligent management method comprising the following steps:
[0005] Step S1: Set initial tags and keywords, label historical data and classify them according to the tags, analyze behavioral patterns, trends, and new attack features, and build an artificial intelligence anomaly feature map model.
[0006] Step S1-1: Use autoencoders to search for images and videos, use simple comprehension networks (RBMs) to process unstructured data, use deep belief networks (DBNs) to process structured data, and automatically identify and label unlabeled data.
[0007] Step S1-2: Use neural networks to conduct deep learning and data mining on problem data, use DNN deep learning and Python data analysis to identify various types of artificial intelligence anomalies, construct abnormal scenarios, perform correlation comparison analysis on data, use causal models to determine the results, and construct an artificial intelligence anomaly feature map model.
[0008] Step S2: Calculate the probability of AI anomaly, formulate AI anomaly determination indicators, determine AI anomaly based on the indicators, eliminate misjudgments based on the spatiotemporal graph convolutional network, and issue an early warning when AI anomaly is determined to have occurred.
[0009] Step S2-1: Collect the output data of the artificial intelligence abnormal feature map, and set the weight ratio by comparing the prediction variance of the model output, the characteristics of the current behavior with the historical normal behavior, and the correlation anomalies between each module;
[0010] Step S2-2: The prediction variance accounts for 30%, the behavioral feature deviation accounts for 40%, and the associated anomaly accounts for 30%. The comprehensive anomaly probability is generated by weighted summation.
[0011] Step S2-3: Set thresholds for normal, suspicious, and abnormal AI behavior, and mark abnormal behavior. Specifically, if the output variance exceeds twice the historical mean for five consecutive times, the behavior pattern deviates from normal behavior characteristics by more than 40%, and there are more than two abnormal nodes in the cross-module call chain;
[0012] Step S2-4, the variance calculation formula is:
[0013]
[0014] Where S var represents the variance score, u var is the historical normal variance mean. If the variance is less than the historical normal variance mean for five consecutive times, the score is 0. If the variance is greater than the normal variance mean for five consecutive times, the score is 1. If it is in the middle, the score is calculated according to the formula. u is the current variance.
[0015] Step S2-5: The characteristic deviation calculation formula is:
[0016]
[0017] Where S dev represents the feature deviation score, d is the deviation, d max is the deviation threshold, which is a min function as a whole, indicating the minimum value of the two values in the brackets;
[0018] Step S2-6, the module association anomaly score calculation formula is:
[0019]
[0020] Where S space represents the module association anomaly score, t x represents the number of abnormal nodes, t represents two thresholds, and the min function takes the minimum value;
[0021] Step S2-7: Calculate the comprehensive abnormality probability by weighted summation.
[0022] Step S2-8: If the abnormality probability is ≥0.6 and ≥2 basic indicators are abnormal, it is determined to be a suspicious state; if the abnormality probability is ≥0.8 and ≥3 basic indicators are abnormal, it is determined to be an abnormal state.
[0023] Step S2-9: Eliminate misjudgment using a spatiotemporal convolutional network;
[0024] Step S2-10: Classify the warning levels, mark yellow warning as suspicious state and red warning as abnormal state.
[0025] Step S3: Calculate the cost after an AI anomaly occurs and set a warning risk level based on the cost.
[0026] Step S3-1: Collect ethical norms and legal rules from different fields, generate a rule base, and clarify the connotation and violation definition of each rule;
[0027] Step S3-2: Integrate and sort out the ethical standards and legal rules of various industries to form a set of rules;
[0028] Step S3-3: For each abnormal feature A of the artificial intelligence abnormal graph i Pre-rule R j Mapping to clarify the relationship;
[0029] Step S3-4: For each abnormal feature A i According to its violation of rule R j severity, assigning a violation score of V ij , determine each rule R j The weight W of the industry j ;
[0030] Step S3-5: For each abnormal feature A i , calculate its risk value R i , the calculation formula is:
[0031]
[0032] Where R i represents the risk value of the i-th abnormal feature, V ij represents the score of the i-th abnormal feature violating the j-th rule, W j Indicates the weight ratio of the jth rule;
[0033] Step S3-6: Calculate the comprehensive risk value for the presence of multiple abnormal features in artificial intelligence. The calculation formula is:
[0034]
[0035] Where R total represents the comprehensive risk value, R i Abnormal characteristic A for destroying a country i Value at risk;
[0036] Step S3-7: Divide the risk level according to the comprehensive risk value, set risk thresholds x1, x2, x3, and x4, and set the risk interval, [x1-x2] for low risk, [x2-x3] for medium risk, and [x3-x4] for high risk;
[0037] Step S3-8: Use the structural causal model to analyze the causal relationship between artificial intelligence behavior and risk consequences, and convert the analysis results of the structural causal model into an intuitive form.
[0038] Step S4: Based on the abnormal feature map, visualize the action chain of AI anomalies, set weights based on the frequency and cost of AI anomalies, predict the AI attack action chain based on the map, and perform different types of defense based on the risk level;
[0039] Step S4-1: Quantify the costs from three aspects: direct losses, indirect losses, and social impacts. The calculation formula is:
[0040]
[0041] Where x represents the cost, W i Represents the weight of each dimension, V i Indicates the quantitative value of each dimension;
[0042] Step S4-2: Set different thresholds s1, s2, and s3 according to the frequency of occurrence of abnormal data in the artificial intelligence abnormality map, representing low frequency, medium frequency, and high frequency respectively, and set different weight ratios according to the frequency of occurrence and cost;
[0043] Step S4-2: Professionals set the abnormality frequency intervals [x1, x2], [x2, x3], and [x3, x4] according to the business scenario. The abnormality frequency in [x1, x2] is low frequency, in [x2, x3] is medium frequency, and in [x3, x4] is high frequency. Different thresholds s1, s2, and s3 are set according to the frequency of abnormal data in the artificial intelligence abnormality map, representing low frequency, medium frequency, and high frequency respectively. Different weight ratios are set according to the frequency of occurrence and cost;
[0044] Step S4-4: When a low-risk anomaly occurs, the protection subsystem is activated to continuously collect anomaly data, record trigger conditions and behavior models, adjust algorithm parameter deviations, and maintain high-frequency monitoring of the anomaly module to ensure recovery to normal;
[0045] Step S4-5: For medium-risk anomalies, activate the protection subsystem to quickly isolate the abnormal module, use redundancy algorithms and backup modules to maintain business operations, and then check the code of the abnormal module, track the abnormal data, and locate the root cause of the abnormal data;
[0046] Step S4-6: alert staff to issue early warnings for high-risk anomalies;
[0047] Step S5: The system updates and adjusts the prediction indicators in real time to cope with the ever-changing artificial intelligence anomalies;
[0048] Step S5-1: Collect historical attack methods and build a historical attack library. Use an anomaly detection algorithm to compare the real-time attack methods with the attack methods in the historical attack library. If the action link, data characteristics, and network request do not match the content in the historical attack library, it is determined to be a new attack;
[0049] Step S5-2: After the defense is completed, the defense method is scored using a quantitative indicator scoring method and a passing score is set. If the score exceeds the passing score, the system automatically extracts the action link, data type, and network request of the new attack, marks the new attack method, and stores it in the historical attack database. If the score does not exceed the passing score, the system does not store the defense method in the historical attack database.
[0050] Step S5-3: Based on the characteristics of the new attack, the protection subsystem is updated in real time by adding new rules to the firewall and updating the rule file to define the pattern matching rules of the new attack characteristics.
[0051] Compared with the prior art, the present invention has the following beneficial effects:
[0052] 1. This invention automatically identifies labels and constructs anomaly maps by fusing images, videos, and structured / unstructured data, thereby improving recognition comprehensiveness and efficiency.
[0053] 2. This invention predicts the entire attack chain and responds according to risk levels, achieving accurate early warning and disposal.
[0054] 3. The present invention responds to abnormal evolution by dynamically updating the model, and the blockchain evidence storage ensures that the data cannot be tampered with, thereby enhancing the system robustness and traceability. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 A schematic flow chart of the steps of an analysis method for preventing AI data anomalies according to the present invention;
[0056] Figure 2 This is a schematic diagram of the workflow of an analysis method for preventing AI data anomalies according to the present invention. DETAILED DESCRIPTION
[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0058] Step S1: Set initial tags and keywords, label historical data and classify them according to the tags, analyze behavioral patterns, trends, and new attack features, and build an artificial intelligence anomaly feature map model.
[0059] Step S1-1: Use autoencoders to search for images and videos, use simple comprehension networks (RBMs) to process unstructured data, use deep belief networks (DBNs) to process structured data, and automatically identify and label unlabeled data.
[0060] Step S1-2: Use neural networks to conduct deep learning and data mining on problem data, use DNN deep learning and Python data analysis to identify various types of artificial intelligence anomalies, construct abnormal scenarios, perform correlation comparison analysis on data, use causal models to determine the results, and construct an artificial intelligence anomaly feature map model.
[0061] Step S2: Calculate the probability of AI anomaly, formulate AI anomaly determination indicators, determine AI anomaly based on the indicators, eliminate misjudgments based on the spatiotemporal graph convolutional network, and issue an early warning when AI anomaly is determined to have occurred.
[0062] Step S2-1: Collect the output data of the artificial intelligence abnormal feature map, and set the weight ratio by comparing the prediction variance of the model output, the characteristics of the current behavior with the historical normal behavior, and the correlation anomalies between each module;
[0063] Step S2-2: The prediction variance accounts for 30%, the behavioral feature deviation accounts for 40%, and the associated anomaly accounts for 30%. The comprehensive anomaly probability is generated by weighted summation.
[0064] Step S2-3: Set thresholds for normal, suspicious, and abnormal AI behavior, and mark abnormal behavior. Specifically, if the output variance exceeds twice the historical mean for five consecutive times, the behavior pattern deviates from normal behavior characteristics by more than 40%, and there are more than two abnormal nodes in the cross-module call chain;
[0065] Step S2-4, the variance calculation formula is:
[0066]
[0067] Where S varrepresents the variance score, u var is the historical normal variance mean. If the variance is less than the historical normal variance mean for five consecutive times, the score is 0. If the variance is greater than the normal variance mean for five consecutive times, the score is 1. If it is in the middle, the score is calculated according to the formula. u is the current variance.
[0068] Step S2-5: The characteristic deviation calculation formula is:
[0069]
[0070] Where S dev represents the feature deviation score, d is the deviation, d max is the deviation threshold, which is a min function as a whole, indicating the minimum value of the two values in the brackets;
[0071] Step S2-6, the module association anomaly score calculation formula is:
[0072]
[0073] Where S space represents the module association anomaly score, t x represents the number of abnormal nodes, t represents two thresholds, and the min function takes the minimum value;
[0074] Step S2-7: Calculate the comprehensive abnormality probability by weighted summation.
[0075] Step S2-8: If the abnormality probability is ≥0.6 and ≥2 basic indicators are abnormal, it is determined to be a suspicious state; if the abnormality probability is ≥0.8 and ≥3 basic indicators are abnormal, it is determined to be an abnormal state.
[0076] Step S2-9: Eliminate misjudgment using a spatiotemporal convolutional network;
[0077] Step S2-10: Classify the warning levels, mark yellow warning as suspicious state and red warning as abnormal state.
[0078] Step S3: Calculate the cost after an AI anomaly occurs and set a warning risk level based on the cost.
[0079] Step S3-1: Collect ethical norms and legal rules from different fields, generate a rule base, and clarify the connotation and violation definition of each rule;
[0080] Step S3-2: Integrate and sort out the ethical standards and legal rules of various industries to form a set of rules;
[0081] Step S3-3: For each abnormal feature A of the artificial intelligence abnormal graph i Pre-rule R j Map and clarify the relationship;
[0082] Step S3-4: For each abnormal feature A i According to its violation of rule R j severity, assigning a violation score of V ij , determine each rule R j The weight W of the industry j ;
[0083] Step S3-5: For each abnormal feature A i , calculate its risk value R i , the calculation formula is:
[0084]
[0085] Where R i represents the risk value of the i-th abnormal feature, V ij represents the score of the i-th abnormal feature violating the j-th rule, W j Indicates the weight ratio of the jth rule;
[0086] Step S3-6: Calculate the comprehensive risk value for the presence of multiple abnormal features in artificial intelligence. The calculation formula is:
[0087]
[0088] Where R total represents the comprehensive risk value, R i Abnormal characteristic A for destroying a country i Value at risk;
[0089] Step S3-7: Divide the risk level according to the comprehensive risk value, set risk thresholds x1, x2, x3, and x4, and set the risk interval, [x1-x2] for low risk, [x2-x3] for medium risk, and [x3-x4] for high risk;
[0090] Step S3-8: Use the structural causal model to analyze the causal relationship between artificial intelligence behavior and risk consequences, and convert the analysis results of the structural causal model into an intuitive form.
[0091] Step S4: Based on the abnormal feature map, visualize the action chain of AI anomalies, set weights based on the frequency and cost of AI anomalies, predict the AI attack action chain based on the map, and perform different types of defense based on the risk level;
[0092] Step S4-1: Quantify the costs from three aspects: direct losses, indirect losses, and social impacts. The calculation formula is:
[0093]
[0094] Where x represents the cost, Wi Represents the weight of each dimension, V i Indicates the quantitative value of each dimension;
[0095] Step S4-2: Professionals set the abnormality frequency intervals [x1, x2], [x2, x3], and [x3, x4] according to the business scenario. The abnormality frequency in [x1, x2] is low frequency, in [x2, x3] is medium frequency, and in [x3, x4] is high frequency. Different thresholds s1, s2, and s3 are set according to the frequency of abnormal data in the artificial intelligence abnormality map, representing low frequency, medium frequency, and high frequency respectively. Different weight ratios are set according to the frequency of occurrence and cost;
[0096] Step S4-3: Utilize the weights of the abnormal feature graph to predict potential AI attack action links through the algorithm model, analyze existing features to infer possible future attack paths and impacts, and identify risks in advance;
[0097] Step S4-4: When a low-risk anomaly occurs, the protection subsystem is activated to continuously collect anomaly data, record trigger conditions and behavior models, adjust algorithm parameter deviations, and maintain high-frequency monitoring of the anomaly module to ensure recovery to normal;
[0098] Step S4-5: For medium-risk anomalies, activate the protection subsystem to quickly isolate the abnormal module, use redundancy algorithms and backup modules to maintain business operations, and then check the code of the abnormal module, track the abnormal data, and locate the root cause of the abnormal data;
[0099] Step S4-6: alert staff to issue early warnings for high-risk anomalies;
[0100] Step S5: The system updates and adjusts the prediction indicators in real time to cope with the ever-changing artificial intelligence anomalies;
[0101] Step S5-1: Collect historical attack methods and build a historical attack library. Use an anomaly detection algorithm to compare the real-time attack methods with the attack methods in the historical attack library. If the action link, data characteristics, and network request do not match the content in the historical attack library, it is determined to be a new attack;
[0102] Step S5-2: After the defense is completed, the defense method is scored using a quantitative indicator scoring method and a passing score is set. If the score exceeds the passing score, the system automatically extracts the action link, data type, and network request of the new attack, marks the new attack method, and stores it in the historical attack database. If the score does not exceed the passing score, the system does not store the defense method in the historical attack database.
[0103] Step S5-3: Based on the characteristics of the new attack, the protection subsystem is updated in real time by adding new rules to the firewall and updating the rule file to define the pattern matching rules of the new attack characteristics.
[0104] Example 1: Use an autoencoder to scan ATM surveillance video and identify abnormal behaviors, such as masked operation and multiple people watching. Use RBM to process customer call voice and extract keywords. Use DBN to process transaction flow and automatically tag labels such as "high-frequency cross-border transfer". Use DNN to analyze 1 million normal transaction data in 1 year and extract features such as transaction time, amount, location, device fingerprint, etc., to construct abnormal scenarios, such as "the same account conducts account transactions in Beijing and New York within 1 hour". Use causal models to determine associations. For example, the probability of triggering an abnormality for the combination of "remote login + large-amount transfer" reaches 90%. Generate an abnormal feature map. When detecting whether the user's third transaction of the day is abnormal, it is necessary to collect the user's historical transaction data, calculate the prediction variance, behavioral feature deviation, and the weight ratio of the model-related abnormality, and set a threshold. The normal threshold is when the variance is less than or equal to the historical mean, and the abnormal threshold is when the variance is greater than twice the historical mean for five consecutive times or the feature deviation is greater than 40%. Call the payment system and use The user authentication system's abnormal node: the variance of the user's third transaction that day was 3-mean, and the variance was greater than the two historical means for five consecutive times. The transaction location frequently changed from Shanghai to Guangzhou, with a deviation of 60%. This user's third transaction triggered simultaneous alerts from the anti-money laundering system and the risk control system. The spatiotemporal graph convolutional network was used to verify that the user had no historical abnormal records. However, the transaction location did not match the IP address, eliminating false positives and triggering a red alert. The account was frozen and manually reviewed. The rule weights were set according to national laws, and the score was calculated based on the violation score, with a score of 59. When the device fingerprint anomaly risk was also present, the fingerprint anomaly risk score was calculated to be 40, and the comprehensive contribution value was 99. The risk thresholds were set at 30 for low risk, 60 for medium risk, 80 for high risk, and 100 for extremely high risk. The calculation cost was 530,000 yuan, and the total risk cost was 636,000 yuan after substituting the weight settings. The attack chain was predicted based on the map, and the defense solution was then activated. The attack chain was stored in the anomaly map, and new attacks were then identified based on the updated map.
[0105] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.
Claims
1. An analysis method for preventing AI data anomalies, characterized by: The intelligent management method comprises the following steps: Step S1: Set initial tags and keywords, label historical data and classify them according to the tags, analyze behavioral patterns, trends, and new attack features, and build an artificial intelligence anomaly feature map model. Step S2: Calculate the probability of AI anomaly, formulate AI anomaly determination indicators, determine AI anomaly based on the indicators, eliminate misjudgments based on the spatiotemporal graph convolutional network, and issue an early warning when AI anomaly is determined to have occurred. Step S3: Calculate the cost after an AI anomaly occurs and set a warning risk level based on the cost. Step S4: Based on the abnormal feature map, visualize the action chain of AI anomalies, set weights based on the frequency and cost of AI anomalies, predict the AI attack action chain based on the map, and perform different types of defense based on the risk level; Step S5: The system updates and adjusts the prediction indicators in real time to cope with the ever-changing artificial intelligence anomalies; 2. The analysis method for preventing AI data anomalies according to claim 1, characterized in that: The specific steps of step S1 are as follows: Step S1-1: Use autoencoders to search for images and videos, use simple comprehension networks (RBMs) to process unstructured data, use deep belief networks (DBNs) to process structured data, and automatically identify and label unlabeled data. Step S1-2: Use neural networks to conduct deep learning and data mining on problem data, use DNN deep learning and Python data analysis to identify various types of artificial intelligence anomalies, construct abnormal scenarios, perform correlation comparison analysis on data, use causal models to determine the results, and construct an artificial intelligence anomaly feature map model.
3. The analysis method for preventing AI data anomalies according to claim 2, characterized in that: The specific steps of probability calculation in step S2 are as follows: Step S2-1: Collect the output data of the artificial intelligence abnormal feature map, and set the weight ratio by comparing the prediction variance of the model output, the characteristics of the current behavior with the historical normal behavior, and the correlation anomalies between each module; Step S2-2: The prediction variance accounts for 30%, the behavioral feature deviation accounts for 40%, and the associated anomaly accounts for 30%. The comprehensive anomaly probability is generated by weighted summation.
4. The analysis method for preventing AI data anomalies according to claim 3, characterized in that: The specific steps of determining the index in step S2 are as follows: Step S2-3: Set thresholds for normal, suspicious, and abnormal AI behavior, and mark abnormal behavior. Specifically, if the output variance exceeds twice the historical mean for five consecutive times, the behavior pattern deviates from normal behavior characteristics by more than 40%, and there are more than two abnormal nodes in the cross-module call chain; Step S2-4, the variance calculation formula is: Where S var represents the variance score, u var is the historical normal variance mean. If the variance is less than the historical normal variance mean for five consecutive times, the score is 0. If the variance is greater than the normal variance mean for five consecutive times, the score is 1. If it is in the middle, the score is calculated according to the formula. u is the current variance. Step S2-5: The characteristic deviation calculation formula is: Where S dev represents the feature deviation score, d is the deviation, d max is the deviation threshold, which is a min function as a whole, indicating the minimum value of the two values in the brackets; Step S2-6, the module association anomaly score calculation formula is: Where S space represents the module association anomaly score, t x represents the number of abnormal nodes, t represents two thresholds, and the min function takes the minimum value; Step S2-7: Calculate the comprehensive abnormality probability by weighted summation. Step S2-8: If the abnormality probability is ≥0.6 and ≥2 basic indicators are abnormal, it is determined to be a suspicious state; if the abnormality probability is ≥0.8 and ≥3 basic indicators are abnormal, it is determined to be an abnormal state.
5. The analysis method for preventing AI data anomalies according to claim 4, characterized in that: The specific steps of the early warning in step S2 are as follows: Step S2-9: Eliminate misjudgment using a spatiotemporal convolutional network; Step S2-10: Classify the warning levels, mark yellow warning as suspicious state and red warning as abnormal state.
6. The analysis method for preventing AI data anomalies according to claim 5, characterized in that: The specific steps of dividing the risk levels in step S3 are as follows: Step S3-1: Collect ethical norms and legal rules from different fields, generate a rule base, and clarify the connotation and violation definition of each rule; Step S3-2: Integrate and sort out the ethical standards and legal rules of various industries to form a set of rules; Step S3-3: For each abnormal feature A of the artificial intelligence abnormal graph i Pre-rule R j Map and clarify the relationship; Step S3-4: For each abnormal feature A i According to its violation of rule R j severity, assigning a violation score of V ij , determine each rule R j The weight W of the industry j ; Step S3-5: For each abnormal feature A i , calculate its risk value R i , the calculation formula is: Where R i represents the risk value of the i-th abnormal feature, V ij represents the score of the i-th abnormal feature violating the j-th rule, W j Indicates the weight ratio of the jth rule; Step S3-6: Calculate the comprehensive risk value for the presence of multiple abnormal features in artificial intelligence. The calculation formula is: Where R total represents the comprehensive risk value, R i Abnormal characteristic A for destroying a country i Value at risk; Step S3-7: Divide the risk level according to the comprehensive risk value, set risk thresholds x1, x2, x3, and x4, and set the risk interval, [x1-x2] for low risk, [x2-x3] for medium risk, and [x3-x4] for high risk; 7. The analysis method for preventing AI data anomalies according to claim 6, characterized in that: The specific steps of risk visualization in step S3 are as follows: Step S3-8: Use the structural causal model to analyze the causal relationship between artificial intelligence behavior and risk consequences, and convert the analysis results of the structural causal model into an intuitive form.
8. The analysis method for preventing AI data anomalies according to claim 7, characterized in that: The specific steps of setting the cost and weight in step S4 are as follows: Step S4-1: Quantify the costs from three aspects: direct losses, indirect losses, and social impacts. The calculation formula is: Where x represents the cost, W i Represents the weight of each dimension, V i Indicates the quantitative value of each dimension; Step S4-2: Professionals set the abnormality frequency intervals [x1, x2], [x2, x3], and [x3, x4] according to the business scenario. The abnormality frequency in [x1, x2] is low frequency, in [x2, x3] is medium frequency, and in [x3, x4] is high frequency. Different thresholds s1, s2, and s3 are set according to the frequency of abnormal data in the artificial intelligence abnormality map, representing low frequency, medium frequency, and high frequency respectively. Different weight ratios are set according to the frequency of occurrence and cost; Step S4-3: Utilize the weights of the abnormal feature graph to predict potential AI attack action links through the algorithm model, analyze existing features to infer possible future attack paths and impacts, and identify risks in advance; 9. The analysis method for preventing AI data anomalies according to claim 8, characterized in that: The specific steps of the defense type in step S4 are as follows: Step S4-4: When a low-risk anomaly occurs, the protection subsystem is activated to continuously collect anomaly data, record trigger conditions and behavior models, adjust algorithm parameter deviations, and maintain high-frequency monitoring of the anomaly module to ensure recovery to normal; Step S4-5: For medium-risk anomalies, activate the protection subsystem to quickly isolate the abnormal module, use redundancy algorithms and backup modules to maintain business operations, and then check the code of the abnormal module, track the abnormal data, and locate the root cause of the abnormal data; Step S4-6: alert staff to issue early warnings for high-risk anomalies; 10. The analysis method for preventing AI data anomalies according to claim 9, characterized in that: The specific steps of step S5 are as follows: Step S5-1: Collect historical attack methods and build a historical attack library. Use an anomaly detection algorithm to compare the real-time attack methods with the attack methods in the historical attack library. If the action link, data characteristics, and network request do not match the content in the historical attack library, it is determined to be a new attack; Step S5-2: After the defense is completed, the defense method is scored using a quantitative indicator scoring method and a passing score is set. If the score exceeds the passing score, the system automatically extracts the action link, data type, and network request of the new attack, marks the new attack method, and stores it in the historical attack database. If the score does not exceed the passing score, the system does not store the defense method in the historical attack database. Step S5-3: Based on the characteristics of the new attack, the protection subsystem is updated in real time by adding new rules to the firewall and updating the rule file to define the pattern matching rules of the new attack characteristics.