A data defense detection method and system based on artificial intelligence

Abstract

The present invention discloses an artificial intelligence-based data defense detection method and system. The artificial intelligence-based data defense detection method comprises: S110, collecting data requiring defense detection as input data; S120, generating a unique dynamic fingerprint for the input data to identify potential attack variants; S130, quantifying the likelihood of the data being tampered with by adversarial samples; S140, calculating a threat fusion score based on the dynamic fingerprint and the likelihood of the data being tampered with by adversarial samples; S150, performing data defense detection and response based on the threat fusion score. This application provides continuously updated threat intelligence for data defense detection, facilitating the system's timely optimization of defense strategies to respond to new attack methods and data tampering techniques.

Description

Technical Field

[0001] The present invention relates to an artificial intelligence data detection method, and in particular to an artificial intelligence-based data defense detection method. Background Art

[0002] With the rapid development of artificial intelligence (AI) and machine learning (ML) technologies, AI has been widely applied in various fields, including data analysis, automated decision-making, and image recognition. However, with the widespread adoption of AI technology, data security issues have become increasingly prominent, particularly in the areas of adversarial attacks and data tampering. Adversarial attacks, through subtle perturbations or carefully crafted data falsification, can mislead AI models into making incorrect decisions, thereby threatening data integrity and system security. These attacks are often difficult to detect by traditional security mechanisms, and the increasing sophistication of attack methods poses significant challenges to existing defense technologies.

[0003] While existing AI-based data defense and detection methods have demonstrated some effectiveness in many areas, they still face numerous technical challenges. First, existing dynamic fingerprinting and behavioral analysis technologies often rely on traditional pattern recognition methods, which are susceptible to interference from noisy data or camouflaged attacks, leading to false positives or false negatives. Attackers can circumvent these detection methods by mimicking normal user behavior, making defense more difficult. Second, while defenses against adversarial examples are continuously evolving, existing techniques such as adversarial training and input data preprocessing often fail to completely eliminate the threat of new adversarial attacks and may even degrade AI model performance. Furthermore, computational overhead and real-time responsiveness during the defense process are weaknesses in current technologies. As attack methods continue to evolve and become more complex, existing defense and detection methods need to be continuously optimized and updated to adapt to the challenges of new threats. Therefore, designing more efficient, accurate, and robust AI data defense and detection methods is crucial to ensuring the security and reliability of AI systems. To address this issue, AI-based data defense and detection methods have emerged. This method combines machine learning, deep learning, and data mining technologies, and uses multi-dimensional security protection strategies such as dynamic fingerprint analysis, behavior monitoring, and adversarial sample detection to monitor and analyze the input data of the AI ​​system in real time in order to identify potential malicious attacks and abnormal behaviors. Summary of the Invention

[0004] The present invention provides a data defense detection method based on artificial intelligence, comprising:

[0005] S110, collecting data required for defense detection as input data;

[0006] S120, generating a unique dynamic fingerprint for the input data to identify potential attack variants;

[0007] S130, the possibility of quantifying data being tampered with by adversarial examples;

[0008] S140. Calculate a threat fusion score based on the dynamic fingerprint and the likelihood that the data has been tampered with by the adversarial sample.

[0009] S150. Perform data defense detection and response based on threat fusion scores.

[0010] In the aforementioned artificial intelligence-based data defense detection method, the method for generating a unique dynamic fingerprint specifically includes the following sub-steps:

[0011] Extract local features from input data and perform weighting and normalization;

[0012] The weighted normalized features are compressed into a stable range through the activation function;

[0013] Generates a unique dynamic fingerprint for the processed input data.

[0014] As described above, a data defense detection method based on artificial intelligence is described, in which a unique dynamic fingerprint represents a unique and robust fingerprint code generated for each piece of input data. This code is used to detect and distinguish different data samples, identify and track potential attack variants. The dynamic fingerprint of the input data is calculated through a differentiable hash network. If the attacker slightly tampers with the input data, the dynamic fingerprint will show obvious differences.

[0015] The above-mentioned AI-based data defense detection method, in which the threat fusion score is calculated based on the dynamic fingerprint and the possibility of data being tampered with by the adversarial sample, is specifically divided into the following sub-steps:

[0016] Calculate dynamic fingerprint deviation;

[0017] The threat fusion score is calculated based on the dynamic fingerprint deviation and the possibility that the data has been tampered with by the adversarial sample.

[0018] As described above, an AI-based data defense detection method is described, in which the threat fusion score represents a quantitative indicator of the threat level of a sample. A larger value indicates that the sample is more likely to have been attacked or tampered with. Dynamic fingerprint deviation and adversarial potential respectively provide different abnormality indicators.

[0019] The present invention also provides an artificial intelligence-based data defense detection system, including: a data acquisition module, a dynamic fingerprint generation module, a possibility quantification module, a threat fusion score calculation module, and a detection and response module.

[0020] Data collection module: collects data required for defense detection as input data;

[0021] Dynamic fingerprint generation module: generates a unique dynamic fingerprint for input data to identify potential attack variants;

[0022] Likelihood Quantification Module: Quantifies the likelihood of data being tampered with by adversarial samples;

[0023] Threat fusion score calculation module: Calculates the threat fusion score based on the dynamic fingerprint and the possibility of data being tampered with by adversarial samples;

[0024] Detection and response module: performs data defense detection and response based on threat fusion scores.

[0025] In the aforementioned artificial intelligence-based data defense detection system, the method for generating a unique dynamic fingerprint specifically includes the following sub-steps:

[0026] Extract local features from input data and perform weighting and normalization;

[0027] The weighted normalized features are compressed into a stable range through the activation function;

[0028] Generates a unique dynamic fingerprint for the processed input data.

[0029] As described above, a data defense detection system based on artificial intelligence is described, in which a unique dynamic fingerprint represents a unique and robust fingerprint code generated for each piece of input data. This code is used to detect and distinguish different data samples, identify and track potential attack variants. The dynamic fingerprint of the input data is calculated through a differentiable hash network. If the attacker slightly tampers with the input data, the dynamic fingerprint will show obvious differences.

[0030] In the aforementioned AI-based data defense detection system, a method for calculating a threat fusion score based on dynamic fingerprints and the likelihood of data being tampered with by adversarial samples is specifically divided into the following sub-steps:

[0031] Calculate dynamic fingerprint deviation;

[0032] The threat fusion score is calculated based on the dynamic fingerprint deviation and the possibility that the data has been tampered with by the adversarial sample.

[0033] As described above, an AI-based data defense detection system is described, wherein the threat fusion score represents a quantitative indicator of the threat level of a sample. A larger value indicates that the sample is more likely to have been attacked or tampered with. Dynamic fingerprint deviation and adversarial potential respectively provide different abnormality indicators.

[0034] The beneficial effects achieved by the present invention are as follows: providing continuously updated threat intelligence for data defense detection, which helps the system to optimize defense strategies in a timely manner to deal with new attack methods and data tampering technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0036] Figure 1 This is a flow chart of an artificial intelligence-based data defense detection method provided in Example 1 of the present application;

[0037] Figure 2 This is a system schematic diagram of a train dynamic characteristics detection under a wind tunnel experiment provided in Example 2 of the present application. DETAILED DESCRIPTION

[0038] The following is a clear and complete description of 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 them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0039] Example 1

[0040] like Figure 1 As shown, the first embodiment of the present application provides a data defense detection method based on artificial intelligence, including:

[0041] Step S110: Collect data that needs to be tested for defense as input data;

[0042] The data sources for defense detection mainly include user behavior data, system log and event data, adversarial sample data, and environmental perception data. User behavior data refers to the dynamic fingerprint data collected by tracking user operation behaviors (such as clicks, inputs, operation time, etc.). This data helps identify whether there are signs of abnormal behavior or attack behavior. System log and event data refers to log information obtained from servers, applications, and network devices. These logs contain system access, error messages, abnormal operations, etc., which can provide strong support for defense detection. Adversarial sample data refers to data samples that may affect AI models obtained by simulating adversarial attacks or collecting malicious samples in the network. This data is used to evaluate the robustness of the model to adversarial samples. Environmental perception data refers to data including device information, IP addresses, geographic locations, etc. This data helps determine whether users or devices are within the normal range of activities, thereby improving the ability to identify abnormal behavior.

[0043] To collect data required for defense detection as input data, first deploy a variety of data collection tools, such as behavioral analysis tools, network traffic monitors, and log analysis systems, to automatically collect data from various data sources. User behavior data is captured through interaction logs on the front-end page or activity records on the back-end server to ensure that each user's operation path and time period are reflected. System logs and event data are configured on network devices and servers to regularly transmit key information to the security analysis platform. Adversarial sample data is collected by simulating different attack scenarios, automatically generating or collecting maliciously tampered samples from the network. All data is stored in an encrypted database to ensure data integrity and confidentiality. After processing, it is passed as input to the defense detection system for analysis and response.

[0044] Step S120: Generate a unique dynamic fingerprint for the input data to identify potential attack variants;

[0045] A unique dynamic fingerprint represents a unique, robust fingerprint code generated for each piece of input data. This code can be used to detect and distinguish different data samples, and identify and track potential attack variants. Attack variants refer to data that has been slightly tampered with or countered by an attack. The dynamic fingerprint of the input data is calculated using a differentiable hash network. If an attacker has slightly tampered with the input data, the dynamic fingerprint will show significant differences. The method for generating a unique dynamic fingerprint specifically includes the following sub-steps:

[0046] Step S121: extract local features from the input data, and perform weighting and normalization processing;

[0047] The input data is divided into n sub-segments, each sub-segment is assigned a learnable weight, the weighted value of each sub-segment is calculated, the local features of each sub-segment are obtained, and the local features of the input data are obtained by weighting. Specifically, the formula is used: Extract local features of input data, where Represents the local features of the input data, MID(*) is the feature extraction function, x represents the input data, Indicates the i-th sub-segment, i ranges from 1 to n, represents the learnable weight of each segment, Indicates the normalization process for the i-th word segment.

[0048] Step S122: compressing the weighted normalized features into a stable range through an activation function;

[0049] The activation function is used to compress the weighted normalized features into a stable range to increase the nonlinear expression capability. The activation function uses the sigmoid function. Specifically, the formula is: represents the compression process, where Represents the result after compressing the feature to a stable range, represents the sigmoid function, Represents the local features of the input data.

[0050] Step S123: Generate a unique dynamic fingerprint for the processed input data;

[0051] By using data perception, the fingerprint is made specific to different inputs, and a unique dynamic fingerprint is generated for the processed input data. Specifically, the formula is used: Generate a unique dynamic fingerprint, where Indicates the unique dynamic fingerprint generated, represents the processed input data, represents the sensitive adjustment matrix related to the input data, Represents element-wise multiplication.

[0052] Step S130: quantify the possibility that the data has been tampered with by the adversarial sample;

[0053] In order to quantify the possibility of adversarial tampering of input data, an adversarial gradient field is introduced. This gradient field can quantify the direction and amplitude of data mutations in the input space, thereby effectively identifying potential attack samples. First, the log-likelihood gradient of the model output with respect to the input is calculated. The log-likelihood gradient reflects the model's sensitivity to prediction changes near the current input and indicates in which direction a small perturbation of the data will lead to the maximum prediction change. In order to avoid over-amplifying small perturbations when the input data is still close to a clean sample, an adaptive adjustment factor based on the distance between the input and the corresponding clean sample is introduced. This factor tends to zero when the input is close to the clean sample, thereby effectively suppressing the gradient strength; and when the distance between the input and the clean sample increases, the factor tends to one, fully exposing potential adversarial tampering. Specifically, the formula is used:

[0054] Calculate the probability that the data has been tampered with by the adversarial sample, where represents the possibility, x represents the input data, represents the original clean sample, represents the prediction model, y represents the prediction result, represents the model confidence, It means to find the gradient of data x.

[0055] Step S140: Calculate a threat fusion score based on the dynamic fingerprint and the likelihood that the data has been tampered with by the adversarial sample;

[0056] The method for calculating the threat fusion score based on the dynamic fingerprint and the possibility of data being tampered with by adversarial samples is divided into the following sub-steps:

[0057] Step S141, calculating the dynamic fingerprint deviation;

[0058] To reflect the degree of deviation of the input sample from the clean sample in the feature space, a unique dynamic fingerprint is generated for the input data x. , compared with the original clean sample The generated original clean sample fingerprint is compared to calculate the deviation value. Specifically, the formula is used: Calculate the dynamic fingerprint deviation, where Indicates the dynamic fingerprint deviation value. The higher the deviation value, the more obvious the fingerprint change is and the higher the suspicion is. x represents the input data. represents the original clean sample, Indicates the unique dynamic fingerprint generated, Represents the original clean sample fingerprint.

[0059] Step S142: Calculate a threat fusion score based on the dynamic fingerprint deviation and the likelihood that the data has been tampered with by the adversarial sample;

[0060] The threat fusion score is a quantitative indicator of the threat level of a sample. A larger value indicates that the sample is more likely to have been attacked or tampered with. Dynamic fingerprint deviation and adversarial potential provide different abnormality indicators. Specifically, the formula is:

[0061] Calculate the threat fusion score, where represents the threat fusion score of the input data x, represents the impact factor of dynamic fingerprint deviation on threat fusion score, Indicates the impact factor of the possibility of data being tampered with by adversarial samples on the threat fusion score, represents the dynamic fingerprint deviation, Indicates the possibility that the data is tampered with by the adversarial sample, Represents the normalization function.

[0062] Step S150: Perform data defense detection and response based on the threat fusion score;

[0063] When the threat fusion score When a set security threshold is exceeded, the system automatically triggers a defense detection mechanism. First, the system analyzes the nature of the threat based on the specific source of the score. If a high score stems from a dynamic fingerprint deviation, the system initiates user behavior analysis to identify and isolate possible anomalous behavior data to prevent further data tampering or abnormal operations. If a high score stems from adversarial sample manipulation, the system assesses data integrity using an adversarial sample detection model and initiates data cleansing and repair procedures to remove potential malicious tampering.

[0064] Once a threat is detected, the system automatically executes response measures, enhancing the robustness of the model, adjusting the algorithm to address new attacks, and restricting access rights for suspicious users. Simultaneously, the system initiates a secondary verification mechanism for manual review and confirmation to minimize false positives. Ultimately, the defense mechanism optimizes its strategies based on continuously updated threat intelligence to address new attack methods and data tampering techniques, forming a dynamic, closed-loop defense system.

[0065] Example 2

[0066] like Figure 2 As shown, the first embodiment of the present application provides an artificial intelligence-based data defense detection system, including:

[0067] Data collection module 21: collects data required for defense detection as input data;

[0068] The data sources for defense detection mainly include user behavior data, system log and event data, adversarial sample data, and environmental perception data. User behavior data refers to the dynamic fingerprint data collected by tracking user operation behaviors (such as clicks, inputs, operation time, etc.). This data helps identify whether there are signs of abnormal behavior or attack behavior. System log and event data refers to log information obtained from servers, applications, and network devices. These logs contain system access, error messages, abnormal operations, etc., which can provide strong support for defense detection. Adversarial sample data refers to data samples that may affect AI models obtained by simulating adversarial attacks or collecting malicious samples in the network. This data is used to evaluate the robustness of the model to adversarial samples. Environmental perception data refers to data including device information, IP addresses, geographic locations, etc. This data helps determine whether users or devices are within the normal range of activities, thereby improving the ability to identify abnormal behavior.

[0069] To collect data required for defense detection as input data, first deploy a variety of data collection tools, such as behavioral analysis tools, network traffic monitors, and log analysis systems, to automatically collect data from various data sources. User behavior data is captured through interaction logs on the front-end page or activity records on the back-end server to ensure that each user's operation path and time period are reflected. System logs and event data are configured on network devices and servers to regularly transmit key information to the security analysis platform. Adversarial sample data is collected by simulating different attack scenarios, automatically generating or collecting maliciously tampered samples from the network. All data is stored in an encrypted database to ensure data integrity and confidentiality. After processing, it is passed as input to the defense detection system for analysis and response.

[0070] Dynamic fingerprint generation module 22: generates a unique dynamic fingerprint for input data to identify potential attack variants;

[0071] A unique dynamic fingerprint represents a unique, robust fingerprint code generated for each piece of input data. This code can be used to detect and distinguish different data samples, and identify and track potential attack variants. Attack variants refer to data that has been slightly tampered with or countered by an attack. The dynamic fingerprint of the input data is calculated using a differentiable hash network. If an attacker has slightly tampered with the input data, the dynamic fingerprint will show significant differences. The method for generating a unique dynamic fingerprint specifically includes the following sub-steps:

[0072] Local feature extraction module: extracts local features from input data and performs weighting and normalization processing;

[0073] The input data is divided into n sub-segments, each sub-segment is assigned a learnable weight, the weighted value of each sub-segment is calculated, the local features of each sub-segment are obtained, and the local features of the input data are obtained by weighting. Specifically, the formula is used: Extract local features of input data, where Represents the local features of the input data, MID(*) is the feature extraction function, x represents the input data, Indicates the i-th sub-segment, i ranges from 1 to n, represents the learnable weight of each segment, Indicates the normalization process for the i-th word segment.

[0074] Activation module: compresses the weighted normalized features into a stable range through the activation function;

[0075] The activation function is used to compress the weighted normalized features into a stable range to increase the nonlinear expression capability. The activation function uses the sigmoid function. Specifically, the formula is: represents the compression process, where Represents the result after compressing the feature to a stable range, represents the sigmoid function, Represents the local features of the input data.

[0076] Dynamic fingerprint generation module: generates a unique dynamic fingerprint for the processed input data;

[0077] By using data perception, the fingerprint is made specific to different inputs, and a unique dynamic fingerprint is generated for the processed input data. Specifically, the formula is used: Generate a unique dynamic fingerprint, where Indicates the unique dynamic fingerprint generated, represents the processed input data, represents the sensitive adjustment matrix related to the input data, Represents element-wise multiplication.

[0078] Likelihood Quantification Module 23: Quantifies the likelihood of data being tampered with by adversarial samples;

[0079] In order to quantify the possibility of adversarial tampering of input data, an adversarial gradient field is introduced. This gradient field can quantify the direction and amplitude of data mutations in the input space, thereby effectively identifying potential attack samples. First, the log-likelihood gradient of the model output with respect to the input is calculated. The log-likelihood gradient reflects the model's sensitivity to prediction changes near the current input and indicates in which direction a small perturbation of the data will lead to the maximum prediction change. In order to avoid over-amplifying small perturbations when the input data is still close to a clean sample, an adaptive adjustment factor based on the distance between the input and the corresponding clean sample is introduced. This factor tends to zero when the input is close to the clean sample, thereby effectively suppressing the gradient strength; and when the distance between the input and the clean sample increases, the factor tends to one, fully exposing potential adversarial tampering. Specifically, the formula is used:

[0080] Calculate the probability that the data has been tampered with by the adversarial sample, where represents the possibility, x represents the input data, represents the original clean sample, represents the prediction model, y represents the prediction result, represents the model confidence, It means to find the gradient of data x.

[0081] Threat fusion score calculation module 24: calculates the threat fusion score based on the dynamic fingerprint and the possibility of data being tampered with by the adversarial sample;

[0082] The method for calculating the threat fusion score based on the dynamic fingerprint and the possibility of data being tampered with by adversarial samples is divided into the following sub-steps:

[0083] Dynamic fingerprint difference calculation module: calculate dynamic fingerprint deviation;

[0084] To reflect the degree of deviation of the input sample from the clean sample in the feature space, a unique dynamic fingerprint is generated for the input data x. , compared with the original clean sample The generated original clean sample fingerprint is compared to calculate the deviation value. Specifically, the formula is used: Calculate the dynamic fingerprint deviation, where Indicates the dynamic fingerprint deviation value. The higher the deviation value, the more obvious the fingerprint change is and the higher the suspicion is. x represents the input data. represents the original clean sample, Indicates the unique dynamic fingerprint generated, Represents the original clean sample fingerprint.

[0085] Threat fusion score calculation module: Calculates the threat fusion score based on the dynamic fingerprint deviation and the possibility of data being tampered with by adversarial samples;

[0086] The threat fusion score is a quantitative indicator of the threat level of a sample. A larger value indicates that the sample is more likely to have been attacked or tampered with. Dynamic fingerprint deviation and adversarial potential provide different abnormality indicators. Specifically, the formula is:

[0087] Calculate the threat fusion score, where represents the threat fusion score of the input data x, represents the impact factor of dynamic fingerprint deviation on threat fusion score, Indicates the impact factor of the possibility of data being tampered with by adversarial samples on the threat fusion score, represents the dynamic fingerprint deviation, Indicates the possibility that the data is tampered with by the adversarial sample, Represents the normalization function.

[0088] Detection and response module 25: performs data defense detection and response based on threat fusion scores;

[0089] When the threat fusion score When a set security threshold is exceeded, the system automatically triggers a defense detection mechanism. First, the system analyzes the nature of the threat based on the specific source of the score. If a high score stems from a dynamic fingerprint deviation, the system initiates user behavior analysis to identify and isolate possible anomalous behavior data to prevent further data tampering or abnormal operations. If a high score stems from adversarial sample manipulation, the system assesses data integrity using an adversarial sample detection model and initiates data cleansing and repair procedures to remove potential malicious tampering.

[0090] Once a threat is detected, the system automatically executes response measures, enhancing the robustness of the model, adjusting the algorithm to address new attacks, and restricting access rights for suspicious users. Simultaneously, the system initiates a secondary verification mechanism for manual review and confirmation to minimize false positives. Ultimately, the defense mechanism optimizes its strategies based on continuously updated threat intelligence to address new attack methods and data tampering techniques, forming a dynamic, closed-loop defense system.

[0091] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A data defense detection method based on artificial intelligence, characterized in that: include: S110, collecting data required for defense detection as input data; S120, generating a unique dynamic fingerprint for the input data to identify potential attack variants; S130. The possibility of tampering with quantified data by adversarial samples is expressed as: Calculate the probability that the data has been tampered with by the adversarial sample, where represents the possibility, x represents the input data, represents the original clean sample, represents the prediction model, y represents the prediction result, represents the model confidence, Indicates finding the gradient of data x; S140. Calculate a threat fusion score based on the dynamic fingerprint and the likelihood that the data has been tampered with by the adversarial sample. S150. Perform data defense detection and response based on threat fusion scores.

2. The data defense detection method based on artificial intelligence according to claim 1, characterized in that: The method for generating a unique dynamic fingerprint includes the following sub-steps: Extract local features from input data and perform weighting and normalization; The weighted normalized features are compressed into a stable range through the activation function; Generates a unique dynamic fingerprint for the processed input data.

3. The data defense detection method based on artificial intelligence according to claim 2, characterized in that: A unique dynamic fingerprint represents a unique and robust fingerprint code generated for each piece of input data. This code is used to detect and distinguish different data samples, identify and track potential attack variants. The dynamic fingerprint of the input data is calculated through a differentiable hash network. If the attacker slightly tampers with the input data, the dynamic fingerprint will show obvious differences.

4. The data defense detection method based on artificial intelligence according to claim 1, characterized in that: The method for calculating the threat fusion score based on the dynamic fingerprint and the possibility of data tampering by adversarial samples is divided into the following sub-steps: Calculate dynamic fingerprint deviation; The threat fusion score is calculated based on the dynamic fingerprint deviation and the possibility that the data has been tampered with by the adversarial sample.

5. The data defense detection method based on artificial intelligence according to claim 4, characterized in that: The threat fusion score is a quantitative indicator of the threat level of a sample. A larger value indicates that the sample is more likely to have been attacked or tampered with. Dynamic fingerprint deviation and adversarial potential provide different abnormality indicators.

6. A data defense detection system based on artificial intelligence, characterized in that: include: Data collection module: collects data required for defense detection as input data; Dynamic fingerprint generation module: generates a unique dynamic fingerprint for input data to identify potential attack variants; Possibility quantification module: quantifies the possibility of data being tampered with by adversarial samples. The quantification formula is expressed as: Calculate the probability that the data has been tampered with by the adversarial sample, where represents the possibility, x represents the input data, represents the original clean sample, represents the prediction model, y represents the prediction result, represents the model confidence, Indicates finding the gradient of data x; Threat fusion score calculation module: Calculates the threat fusion score based on the dynamic fingerprint and the possibility of data being tampered with by adversarial samples; Detection and response module: performs data defense detection and response based on threat fusion scores.

7. The artificial intelligence-based data defense detection system according to claim 6, characterized in that: The method for generating a unique dynamic fingerprint includes the following sub-steps: Extract local features from input data and perform weighting and normalization; The weighted normalized features are compressed into a stable range through the activation function; Generates a unique dynamic fingerprint for the processed input data.

8. The artificial intelligence-based data defense detection system according to claim 7, characterized in that: A unique dynamic fingerprint represents a unique and robust fingerprint code generated for each piece of input data. This code is used to detect and distinguish different data samples, identify and track potential attack variants. The dynamic fingerprint of the input data is calculated through a differentiable hash network. If the attacker slightly tampers with the input data, the dynamic fingerprint will show obvious differences.

9. The artificial intelligence-based data defense detection system according to claim 6, characterized in that: The method for calculating the threat fusion score based on the dynamic fingerprint and the possibility of data tampering by adversarial samples is divided into the following sub-steps: Calculate dynamic fingerprint deviation; The threat fusion score is calculated based on the dynamic fingerprint deviation and the possibility that the data has been tampered with by the adversarial sample.

10. The artificial intelligence-based data defense detection system according to claim 9, characterized in that: The threat fusion score is a quantitative indicator of the threat level of a sample. A larger value indicates that the sample is more likely to have been attacked or tampered with. Dynamic fingerprint deviation and adversarial potential provide different abnormality indicators.

Images