An ai logic reasoning security protection method with double entropy value threshold linkage

Abstract

This invention discloses a security protection method for AI logical reasoning based on dual entropy threshold linkage, belonging to the field of artificial intelligence security protection technology. Based on a technical architecture of binary result judgment, dual entropy threshold linkage, low-level interception of non-logical data, and isolation between logic and computation units, this invention constructs a two-dimensional monitoring system including parameter call complex entropy and real-time bias entropy. It accurately judges abnormal behavior of the logic-side AI by setting preset security thresholds and executes graded reset protection operations based on the judgment results. This invention employs a physical layer / protocol layer non-logical data flow interception mechanism to block external injection risks at the source; the protection algorithm does not perform redundant mathematical operations, only undertaking verification and isolation operations, resulting in low computational overhead and fast response speed; the graded reset mechanism can adapt to different security level scenarios, effectively preventing security risks such as abnormal parameter calls, bias drift, and external data injection, filling the technical gap of single-dimensional judgment protection, and significantly improving the security and stability of the AI ​​logical reasoning process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence logic security protection technology, specifically to an AI logic reasoning security protection method based on dual entropy threshold linkage. Background Technology

[0002] With the deep application of artificial intelligence in core scenarios such as logical reasoning and decision-making, problems such as bias drift, abnormal parameter calls, and intrusion of malicious external data streams in the AI ​​logic layer are becoming increasingly prominent. Existing protection solutions mostly focus on encryption protection at the model and data layers, lacking a dual-dimensional entropy monitoring and threshold-linked judgment mechanism for the logical reasoning kernel. Conventional protection either has a single judgment dimension or contains intermediate ambiguous output states, making it easy to bypass and resulting in a high false positive rate, failing to achieve physical isolation protection between the logic side and the computing unit. Existing technologies lack a systematic protection architecture based on dual entropy constraints, binary result judgment, and layered isolation reset, making it difficult to meet the usage requirements of highly reliable AI logic security. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a security protection method for AI logic reasoning based on dual entropy threshold linkage. By constructing a dual entropy threshold system, binary result output rules, data flow layered interception, and a graded reset mechanism, it can achieve accurate identification, rapid judgment, and secure isolation of abnormal AI behavior on the logic side. This invention discloses a security protection method for AI logic reasoning based on dual entropy threshold linkage. Its features include five underlying technical constraint rules: Rule 1: The final output of the logic-side AI is a binary result system with no intermediate ambiguity, outputting only two results: correct or incorrect; Rule 2: If the real-time bias entropy value E_L of the logic-side AI exceeds a preset security threshold, it is determined to be an abnormal output error; Rule 3: The protection algorithm only receives monitoring data output by the logic-side AI; all data streams not from the logic side are directly discarded at the physical or protocol layer and do not participate in any verification logic; Rule 4: The protection algorithm does not perform redundant and complex mathematical operations, only undertaking the two core operations of logic verification and isolation control; Rule 5: The logic-side AI and the associated computational unit achieve complete architectural isolation; the abnormal state and computation results of the associated computational unit do not affect the normal reasoning of the logic-side AI or the judgment logic of the protection algorithm. The protection method of this invention includes the following core steps: Step 1: Obtain the core parameters of the AI ​​on the logic side after desensitization. The core parameters include the weight factor W, the dual entropy value security threshold P0, and E_L0; Step 2: Calculate the first entropy value P through the abnormal parameter call complex formula. The calculation formula is: P = (number of abnormal parameter calls / total number of parameter calls) × weight factor W. The number of abnormal parameter calls is customized by the manufacturer according to its own logic side AI parameter system; Step 3: Collect the real-time bias value of the logic side AI and calculate the second entropy value E_L to establish a real-time entropy value monitoring benchmark; Step 4: Output the judgment result S using the dual entropy value threshold linkage judgment rule; when P ≤ P0 and E_L ≤ E_L0, S = 1; when P > P0 or E_L > E_L0, S = 0; Step 5: Execute the corresponding protection operation according to the judgment result S; when S = 0, clear the current inference thread and reset the associated operation unit in stages; when S = 1, maintain the normal logical inference operation state of the AI. Furthermore, the weighting factor W ranges from 0.5 to 1.0 and can be flexibly set according to the importance of business parameters. Furthermore, the number of abnormal parameter calls includes at least one of the following: parameter call errors, unauthorized calls, and unresponsive calls. Furthermore, the resetting of the associated computation unit is divided into three levels: a light reset only clears the current inference thread cache; a moderate reset clears all caches and temporary data; and a heavy reset fully restores the configuration initialization. Furthermore, the dual-entropy threshold P0 ranges from 0.2 to 0.5, and E_L0 ranges from 0.5 to 0.8, and can be calibrated and adapted according to the application scenario.Furthermore, parameter precision control ensures that the P value retains multiple decimal places, while P0 and E_L0 retain two decimal places. Additionally, the E_L entropy value is monitored in real-time to construct a dynamic smoothing baseline curve for parameter call frequency. If the entropy value fluctuation exceeds a preset range, a sudden change is identified, triggering security protection actions. Beneficial effects

[0004] This invention adopts a dual-entropy threshold linkage architecture, filling the technical gap of single-dimensional judgment; it adopts binary unambiguous output to eliminate bypass space from the logical root; it intercepts non-logical data streams at the bottom layer to block intrusion risks at the source; the architecture isolation design does not interfere with each other, with low protection overhead and fast response; it is hierarchical reset to adapt to different security level scenarios, with strong overall originality, accurate judgment and high stability. Detailed Implementation

[0005] The present invention will be further described in detail below with reference to specific embodiments. The present invention is not limited to the following embodiments. Embodiment 1: Logic-side AI Protection in Financial Risk Control Scenarios In this embodiment, the logic-side AI is used for credit risk control decision-making, and the output result is "pass" or "reject". The protection algorithm of the present invention is adopted, and the steps are as follows: Parameter configuration: weight factor W is set to 0.8, security threshold P0 is set to 0.3, and E_L0 is set to 0.6; the number of abnormal parameter calls is defined as "number of parameter call errors + number of unauthorized calls". P value calculation: the number of parameter call errors in the current session is 5, the number of unauthorized calls is 0, and the total number of parameter calls is 100. The calculated P = (5 / 100) × 0.8 = 0.04. E_L value calculation: Based on the real-time parameter call success rate p=0.95, bias coefficient ρ=0.5, error coefficient ε=0.1, and error term Σ|εᵢ|=0.2, E_L=0.5×log (0.95)+0.1×(1+0.2)≈0.10 is calculated. Judgment and execution: P=0.04≤0.3 and E_L=0.10≤0.6, the judgment result S=1, no operation is performed, and the risk control AI is maintained to operate normally. Example 2: Anomaly protection in industrial control scenarios In this example, the logic-side AI is used for generating industrial equipment control instructions. The protection algorithm of this invention is adopted, and the steps are as follows: Parameter configuration: weight factor W is set to 1.0, safety threshold P0 is set to 0.2, and E_L0 is set to 0.5; the number of abnormal parameter calls is defined as "number of parameter call errors + number of unresponsive calls". P-value calculation: In the current session, there are 3 parameter call errors, 4 non-response instances, and a total of 20 parameter call instances. The calculated P = (7 / 20) × 1.0 = 0.35. Judgment and execution: P = 0.35 > 0.2, judgment result S = 0, triggering a moderate reset operation: clearing the current inference thread cache, resetting the cache and temporary data of the associated computational unit, blocking the execution of abnormal instructions, while not affecting other normal inference threads of the device. Example 3: Dynamic benchmark monitoring scenario In this example, the logic-side AI is used for high-frequency trading decisions. The E_L value is monitored using a dynamic smoothing benchmark curve: a dynamic smoothing benchmark curve is constructed based on historical parameter call frequencies, with a normal fluctuation range of ±10%; when a sudden increase of 30% in the parameter call frequency is detected at a certain moment, exceeding the fluctuation range, it is determined that the E_L value has abruptly changed, directly triggering the judgment result S = 0; a mild reset operation is performed: only the current inference thread cache is cleared, quickly blocking the generation of abnormal trading instructions and protecting trading security.The above embodiments are merely exemplary descriptions of the present invention. Those skilled in the art can adjust the parameter values, reset levels, and exception definitions according to actual application scenarios, all of which fall within the protection scope of the present invention.

[0006] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any equivalent changes, modifications, and optimizations made based on the technical concept and entropy logic architecture of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A logic-side AI security protection algorithm, characterized in that, The algorithm is constructed based on the following five underlying technical constraints: Rule 1: The final output of the logic-side AI is a binary result system with no intermediate ambiguity, outputting only "correct" or "incorrect" results; Rule 2: If the real-time bias (E_L) of the logic-side AI exceeds a preset security threshold, it is judged as an output error; Rule 3: The protection algorithm only receives monitoring data output by the logic-side AI. All data streams from sources other than the logic-side are directly discarded at the physical or protocol layer and do not participate in any verification logic; Rule 4: The protection algorithm does not perform any mathematical operations, but only undertakes two core operations: verification logic and isolation control; Rule 5: The logic-side AI and the associated operation unit can be completely isolated. The abnormal state and operation results of the associated operation unit do not affect the normal reasoning of the logic-side AI and the judgment logic of the protection algorithm; The algorithm includes the following core steps: Step 1: Obtain the core parameters of the logic-side AI after desensitization. The core parameters include the weight factor (W) and the security threshold (P0, E_L0); Step 2: Calculate the P value by calling the complex calculation formula through abnormal parameters. The calculation formula is: P = (Number of abnormal parameter calls / Total number of parameter calls) × Weighting factor (W), where the number of abnormal parameter calls is defined by the manufacturer according to its own logic-side AI parameter system; Step 3: Obtain the real-time bias value (E_L) of the logic-side AI and calculate it using the real-time bias value calculation formula, which is: E_L = ρ·log p + ε·(1 + Σ|εᵢ|), where ρ is the bias value coefficient, p is the parameter call success rate, ε is the error coefficient, and εᵢ is the error term for each parameter call; Step 4: Output the judgment result S using the double threshold verification judgment formula, which is: S=1, when (P≤P0) ∩ (E_L≤E_L0); S=0, when (P>P0) ∪ (E_L>E_L0).

2. The logic-side AI security protection algorithm according to claim 1, wherein, The preferred value range for the weighting factor W is 0.5-1.0, and the manufacturer can set the specific value according to the importance of the parameter.

3. The logic-side AI security protection algorithm according to claim 1, wherein, The definition of the number of abnormal parameter calls includes at least one of the following: the number of calls where parameter calls resulted in errors, the number of calls where parameter calls were unauthorized, and the number of calls where parameter calls did not respond.

4. The logic-side AI security protection algorithm according to claim 1, wherein, The operation of resetting the associated computing unit includes three levels, which manufacturers can choose according to their own architecture and the severity of the anomaly: mild reset (only clears the current inference thread cache), moderate reset (resets all caches, clears caches and temporary data), and severe reset (restores factory settings, clears all configurations and data).

5. The logic-side AI security protection algorithm according to claim 1, characterized in that, The recommended value range for the safety threshold P0 is 0.2-0.5, and the recommended value range for E_L0 is 0.5-0.

8. Manufacturers can calibrate and adjust the specific values ​​according to their own scenarios.

6. The logic-side AI security protection algorithm according to claim 1, characterized in that, The accuracy of the core parameters meets the following requirements: the P value retains n decimal places, where n is a positive integer; P0 and E_L0 retain two decimal places, ensuring a balance between protection accuracy and confidentiality protection.

7. The logic-side AI security protection algorithm according to claim 1, characterized in that, The real-time monitoring of the E_L value further includes: a dynamic smoothing baseline curve based on the AI ​​parameter call frequency on the logic side. When a peak or valley value exceeding the preset fluctuation range is detected in the parameter call, it is determined to be a sudden change, triggering the determination result of S=0.