An automated evaluation and adversarial sample generation method for an AI risk control model

CN122654580APending Publication Date: 2026-08-28SHANGHAI ZHOUEN INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610536720.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-04-22
Publication Date
2026-08-28

AI Technical Summary

Technical Problem

[0003]在相关技术中,对于风控引擎效能的测评与检验逻辑,主要依赖于对静态留存数据进行随机切分,构建独立的测试集与验证集;技术人员通过对比测试集中各样本的真实标签与模型输出的预测标签,计算出精确率、召回率、受试者工作特征曲线下的面积以及柯尔莫哥洛夫斯米尔诺夫统计量等宏观评价指标;当上述统计学指标满足设定的数值门槛时,即判定该风控引擎具备上线部署的标准;然而,上述基于历史静态数据堆叠的测评机制无法真实反映业务现场由于恶意交互产生的动态特征畸变;恶意攻击者往往在业务请求中注入微小的、不易察觉的异常数据,相关测评机制由于缺乏主动探测并描绘高维特征空间中非线性决策边界形态的能力,导致模型在面临此类微小畸变时易发生断崖式的判别失效;这种静态评估逻辑与动态异常环境之间的脱节,使得风控引擎内部的结构性脆弱点被彻底掩盖,难以在业务损失发生前被有效隔离和熔断

Benefits of technology

本发明通过针对风控特征的连续、离散维度差异,分别基于雅可比梯度信息梯度寻优生成连续扰动分量、依托预构建状态转移概率矩阵马尔可夫演化生成离散扰动分量,解决了传统对抗样本生成未区分风控特征类型、扰动生成缺乏针对性的问题,带来了对不同类型风控特征的精准差异化扰动设计效果,实现了对抗样本生成与风控特征数据特性的高度适配,提升了对抗样本的有效性;

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122654580A_ABST
    Figure CN122654580A_ABST
Patent Text Reader

Abstract

The application discloses an automatic evaluation and adversarial sample generation method for an AI risk control model, relates to the technical field of machine learning, and comprises the following steps: obtaining a target risk control model and a basic risk control feature set; generating a discrete disturbance component by Markov evolution of a discrete feature dimension vector based on a pre-constructed state transition probability matrix; generating a corresponding adversarial sample set; inputting the basic risk control feature set and the adversarial sample set into the target risk control model respectively, constructing a confidence decay surface along a disturbance direction based on a predicted probability feature difference value; converting the critical disturbance threshold into a dynamic regularization penalty term and injecting the dynamic regularization penalty term into a loss function of the target risk control model, updating network weights of the target risk control model through back propagation, and completing automatic reinforcement of the target risk control model; and the application realizes automatic mining and directional reinforcement of model structural vulnerability, and improves the model's ability to resist malicious disturbance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to an automated evaluation and adversarial example generation method for AI risk control models. Background Technology

[0002] In the fields of digital finance and online transactions, risk control systems serve as the core hub for intercepting fraudulent transactions and identifying false identities, and are widely responsible for the real-time analysis of massive transaction data. With the expansion of data scale, the introduction of machine learning models as risk control prediction engines has become the basic application architecture in this field. Such models usually rely on historically accumulated static business ledgers and user transaction records as basic datasets. By mining the mapping relationship between data features and labels such as overdue and fraud, a decision-making mechanism is established, and then risk scoring and classification are performed on newly connected business requests.

[0003] In related technologies, the evaluation and verification logic for the performance of risk control engines mainly relies on randomly splitting static retained data to construct independent test and validation sets. Technicians calculate macro-evaluation indicators such as precision, recall, area under the receiver operating characteristic (ROC) curve, and Kolmogorov-Smirnov statistic by comparing the real labels of each sample in the test set with the predicted labels output by the model. When the above statistical indicators meet the set numerical thresholds, the risk control engine is deemed to meet the standards for online deployment. However, the above evaluation mechanism based on historical static data stacking cannot truly reflect the dynamic feature distortions caused by malicious interactions in the business environment. Malicious attackers often inject tiny, imperceptible abnormal data into business requests. Because the relevant evaluation mechanisms lack the ability to actively detect and depict the nonlinear decision boundary morphology in high-dimensional feature space, the model is prone to precipitous judgment failure when faced with such tiny distortions. This disconnect between static evaluation logic and dynamic abnormal environment completely masks the structural vulnerabilities within the risk control engine, making it difficult to effectively isolate and circuit breakers before business losses occur. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides an automated evaluation and adversarial sample generation method for AI risk control models. This technical solution resolves the problems mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An automated evaluation and adversarial example generation method for AI risk control models includes: Obtain the target risk control model and the basic risk control feature set, wherein the basic risk control feature set includes continuous feature dimension vectors and discrete feature dimension vectors; Jacobian gradient information is extracted from the parameter matrix of the target risk control model. Based on the Jacobian gradient information, gradient optimization is performed on the continuous feature dimension vector to generate continuous perturbation components. Based on the pre-constructed state transition probability matrix, Markov evolution is performed on the discrete feature dimension vector to generate discrete perturbation components. The continuous perturbation component and the discrete perturbation component are spliced ​​and fused, and filtered according to preset physical business logic constraints to generate a corresponding set of adversarial samples. The basic risk control feature set and the adversarial sample set are respectively input into the target risk control model to obtain the difference in their predicted probability features. Based on the difference in their predicted probability features, a confidence decay surface along the disturbance direction is constructed. The critical perturbation threshold is extracted based on the confidence decay surface, and the critical perturbation threshold is transformed into a dynamic regularization penalty term and injected into the loss function of the target risk control model. The network weights of the target risk control model are updated through backpropagation to complete the automatic reinforcement of the target risk control model.

[0006] Optionally, generating continuous perturbation components by performing gradient optimization on the continuous feature dimension vector based on the Jacobian gradient information includes: Extract the loss gradient vector of the target risk control model for the target risk category; Calculate the partial derivative matrix of the loss gradient vector in the feature space corresponding to the continuous feature dimension vector; The gradient ascent direction is determined based on the partial derivative matrix, and numerical accumulation is performed in the gradient ascent direction according to a preset iteration step size to generate continuous perturbation components.

[0007] Optionally, based on a pre-constructed state transition probability matrix, Markov evolution is performed on the discrete feature dimension vector to generate discrete perturbation components, including: Read the initial state category node of the discrete feature dimension vector; In the state transition probability matrix, query the transition probability value of each candidate state node corresponding to the initial state category node; The candidate state node with the highest transition probability value and which causes the risk score of the target risk control model to decrease by more than a preset decrease threshold is extracted as the target state, and the discrete disturbance component is generated based on the target state.

[0008] Optionally, filtering can be performed based on preset physical business logic constraints, including: Extract the value range boundary interval of the feature field corresponding to the continuous disturbance component; Determine whether the continuous disturbance component exceeds the value range boundary interval; In cases where the value exceeds the boundary interval of the range, the continuous disturbance component is forcibly mapped to the extreme point of the boundary interval of the range using a piecewise truncation function.

[0009] Optionally, a confidence decay surface along the perturbation direction is constructed based on the predicted probability feature difference, including: The perturbation amplitude in the adversarial sample set is discretized at multiple levels to generate multiple asymptotic perturbation scale vectors. The multiple progressive perturbation scale vectors are sequentially injected into the basic risk control feature set to obtain test sample matrices with different step dimensions; Record the output probability vector of the target risk control model when processing test sample matrices of various step dimensions, use the disturbance amplitude as the independent variable and the output probability vector as the dependent variable to perform polynomial fitting, and generate the confidence decay surface.

[0010] Optionally, the critical perturbation threshold is converted into a dynamic regularization penalty term and injected into the loss function of the target risk control model, including: Calculate the second derivative feature of the confidence decay surface at the critical perturbation threshold, where the second derivative feature characterizes the surface curvature; The surface curvature is linearly weighted with the original cross-entropy loss calculation term, and the weighted product result is extracted to construct the dynamic regularization penalty term. The dynamic regularization penalty term is appended to the end of the original loss function of the target risk control model to form a reconstructed loss function.

[0011] Optionally, before updating the network weights of the target risk control model via backpropagation, the method further includes: Extract the forward propagation intermediate tensor representation of the hidden layers within the target risk control model; The forward propagation intermediate tensor is calculated as the cosine distance vector between the input basic risk control feature set and the input adversarial sample set; The elements in the cosine distance vector that are less than the distance-sensitive threshold are indexed and marked as non-sensitive channels.

[0012] Optionally, the network weights of the target risk control model are updated via backpropagation, including: A binary gradient mask matrix is ​​generated based on the non-sensitive channel; When calculating the gradient tensor of the reconstruction loss function relative to the hidden layer weight matrix, the gradient tensor is multiplied by the gradient mask matrix to generate a modified gradient tensor. The network weights are updated in the opposite direction of the corrected gradient tensor.

[0013] Optionally, after inputting the basic risk control feature set and the adversarial example set into the target risk control model, the method further includes: The basic accuracy features output by the target risk control model based on the basic risk control feature set and the adversarial accuracy features output by the adversarial sample set are obtained respectively. Calculate the ratio of the base accuracy feature to the adversarial accuracy feature, and use the ratio as the adversarial sensitivity weight; The initial Kolmogorov-Smirnov statistic, calculated based on the empirical cumulative distribution function, is feature-scaled using the adversarial sensitivity weights to generate automated evaluation indicators.

[0014] Optionally, the process of updating the network weights of the target risk control model through backpropagation also includes: The norm of the weight difference between two adjacent updates is calculated according to a preset cycle. If the weight difference norm is less than the weight convergence threshold within multiple consecutive cycles, the backpropagation update link is truncated, and the current network weight is frozen.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the problems of traditional adversarial example generation failing to distinguish between risk control feature types and lacking specificity in perturbation generation by generating continuous perturbation components based on Jacobi gradient information gradient optimization and discrete perturbation components based on pre-constructed state transition probability matrices, respectively, targeting the differences in continuous and discrete dimensions of risk control features. It brings about precise and differentiated perturbation design effects for different types of risk control features, achieves a high degree of adaptation between adversarial example generation and risk control feature data characteristics, and improves the effectiveness of adversarial examples. This invention filters the fused perturbation components by combining preset physical business logic constraints and segments and maps continuous perturbation components that exceed the value range boundary. This solves the problems of traditional adversarial sample generation being easily detached from actual risk control business scenarios and generating samples with logical distortion. It brings about a consistent matching effect between adversarial samples and financial risk control physical business logic, and realizes that the generated adversarial sample set fits the real business scenario, avoiding interference from invalid perturbations on model evaluation. This invention obtains the difference in predicted probability features by inputting the basic risk control feature set and the adversarial sample set into the target risk control model. Based on the difference, a confidence decay surface along the perturbation direction is constructed and the critical perturbation threshold is extracted. This solves the problems of not being able to accurately characterize the nonlinear decision boundary shape of AI risk control models in high-dimensional feature space and the lack of accurate quantitative perturbation threshold reference for model hardening. It brings an accurate quantitative characterization effect on the vulnerability of the decision boundary of the risk control model, and realizes the provision of objective and accurate critical perturbation basis for model hardening, and accurately locates the weak boundary of the model's resistance to perturbation. This invention transforms the critical perturbation threshold into a dynamic regularization penalty term and injects it into the loss function of the target risk control model. It also constructs this penalty term by combining the second derivative features of the confidence decay surface. This solves the problems of fixed values, subjective settings, and mismatch between the regularization penalty term of traditional risk control model reinforcement and the actual perturbation resistance capability of the model. It brings about a dynamic adaptation effect between the regularization penalty term and the risk control business scenario and perturbation resistance characteristics of the model, and realizes the accurate reconstruction of the model loss function, making model reinforcement more targeted. This invention extracts the forward propagation intermediate tensor of the model's hidden layers, calculates the cosine distance between the basic features and the adversarial sample input to mark the insensitive channels, and generates a gradient mask matrix to correct the gradient tensor of the backpropagation. This solves the problems of indiscriminate gradient updates and redundant updates of insensitive channels during model hardening, which lead to low hardening efficiency and easy introduction of invalid noise. It brings a directional mask and accurate screening effect for model gradient updates, realizes directional and effective gradient updates, reduces invalid computational overhead, and improves the efficiency and accuracy of model hardening. This invention calculates the ratio of the model's basic accuracy on the basic risk control feature set to the adversarial accuracy on the adversarial sample set, and uses this ratio as the initial KS statistic for adversarial sensitivity weight scaling to generate automated evaluation indicators. This solves the problems of traditional AI risk control model evaluation relying on static historical data, having a single evaluation indicator, and being unable to quantify the model's ability to resist adversarial disturbances. It brings a quantitative evaluation effect of model effectiveness that fits dynamic adversarial scenarios, realizes the construction of a multi-dimensional automated evaluation system, accurately quantifies the adversarial sensitivity of the model, and effectively uncovers the structural vulnerabilities within the model. This invention solves the problems of overfitting caused by excessive weight updates and the lack of clear convergence criteria in traditional model hardening by calculating the norm of the difference between adjacent weights periodically during the backpropagation update process of model weights, and truncating the update link and freezing the weights when the norm of consecutive periods is less than the convergence threshold. It brings the effect of dynamic monitoring and precise control of the model weight update process, avoids the overfitting problem after model hardening, and ensures that the model retains good generalization ability while having the ability to resist disturbances. This invention integrates adversarial sample generation, automated model evaluation, and dynamic robustness reinforcement into a complete end-to-end process. The entire process requires no manual intervention to set parameters and thresholds, solving the problems of traditional AI risk control models being disconnected from evaluation and reinforcement, highly dependent on manual operation, inefficient, and unable to cope with dynamic malicious feature distortions in the business environment. It brings a fully automated processing effect to the AI ​​risk control model from adversarial capability evaluation to robustness reinforcement, significantly reducing the manual cost of model evaluation and reinforcement, effectively making up for the disconnect between static evaluation and dynamic abnormal business environments, and avoiding the model from precipitous judgment failure when facing minor feature distortions. This invention generates an asymptotic perturbation scale vector and constructs a test sample matrix by performing multi-level discretization on the perturbation amplitude of adversarial examples, and fits the confidence decay surface to obtain the result. This solves the problem of not being able to accurately capture the change law of the confidence of the risk control model with feature perturbation, and brings a continuous and visual characterization of the change trend of the model's anti-perturbation ability. It also enables the accurate location of the key perturbation nodes of the model's confidence decay, and provides a refined quantitative reference for the iterative optimization of the model. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 As shown, an automated evaluation and adversarial example generation method for AI risk control models includes: S101, Obtain the target risk control model and the basic risk control feature set, wherein the basic risk control feature set includes a continuous feature dimension vector and a discrete feature dimension vector; S102, extract Jacobian gradient information based on the parameter matrix of the target risk control model, perform gradient optimization on the continuous feature dimension vector based on the Jacobian gradient information to generate continuous perturbation components, and perform Markov evolution on the discrete feature dimension vector based on the pre-constructed state transition probability matrix to generate discrete perturbation components. S103, the continuous perturbation component and the discrete perturbation component are spliced ​​and fused, and filtered according to the preset physical business logic constraints to generate the corresponding adversarial sample set; S104, Input the basic risk control feature set and the adversarial sample set into the target risk control model respectively, obtain the difference in predicted probability features between the two, and construct a confidence decay surface along the disturbance direction based on the difference in predicted probability features; S105, extract the critical perturbation threshold based on the confidence decay surface, convert the critical perturbation threshold into a dynamic regularization penalty term and inject it into the loss function of the target risk control model, update the network weights of the target risk control model through backpropagation, and complete the automatic reinforcement of the target risk control model.

[0019] The above solution achieves end-to-end integration of adversarial sample generation, automated model evaluation, and dynamic robustness reinforcement. It designs differentiated perturbation generation logic based on the continuous and discrete heterogeneous characteristics of risk control features, solving the problems of traditional methods lacking perturbation specificity, adversarial samples being detached from business scenarios, and model reinforcement lacking accurate basis. It effectively uncovers the structural vulnerabilities of AI risk control models and improves the model's anti-perturbation ability and generalization performance in dynamic malicious attack scenarios.

[0020] In some embodiments, obtaining the target risk control model and the basic risk control feature set includes: extracting the deep feedforward network structure parameters and historical baseline sample data deployed in the server memory; parsing the historical baseline sample data into continuous feature dimension vectors and discrete feature dimension vectors; wherein, the continuous feature dimension vectors represent physical quantities with continuous numerical attributes, and the discrete feature dimension vectors represent state codes of enumeration types.

[0021] Specifically, an independent feature registry is established for the feature space differences of heterogeneous data. The dimensional attribute identifier of the corresponding feature is written into the feature registry to ensure the isolation of data paths of different physical attributes in the subsequent mapping logic and prevent computational overflow under non-homogeneous computation graphs. Specifically, the continuous feature dimension vectors after parsing are standardized to map the values ​​to a fixed interval, and the discrete feature dimension vectors are one-hot encoded or embedded vector transformed to ensure that the two types of features have uniform dimensionality adaptability when input into the model, while retaining the original physical meaning and discriminative power of each feature.

[0022] In some embodiments, Jacobian gradient information is extracted based on the parameter matrix of the target risk control model, and gradient optimization is performed on the continuous feature dimension vector based on the Jacobian gradient information to generate continuous perturbation components. Then, Markov evolution is performed on the discrete feature dimension vector based on a pre-constructed state transition probability matrix to generate discrete perturbation components. This includes: constructing a first operational path for continuous features and a second operational path for discrete features; performing forward differential calculation in the first operational path; and performing table lookup and matching in the second operational path. Specifically, the loss gradient vector of the target risk control model for the target risk category is extracted, and the partial derivative matrix of the loss gradient vector in the feature space corresponding to the continuous feature dimension vector is calculated. The partial derivative matrix explicitly indicates the direction of the most drastic numerical increase that leads to the most dramatic decrease in the model output confidence.

[0023] Specifically, the gradient ascent direction is determined based on the partial derivative matrix, and continuous perturbation components are generated by numerical accumulation in the gradient ascent direction according to a preset iteration step size; for discrete features, the initial state category nodes of the discrete feature dimension vector are read, the transition probability values ​​of each candidate state node are queried in the pre-constructed state transition probability matrix, and the node that causes the output to decrease and has the largest probability value is extracted as the target state to generate discrete perturbation components. Specifically, the state transition probability matrix is ​​constructed offline based on historical risk control data. It statistically analyzes the natural transition frequency between each state of discrete features, such as the historical probability of a user's career changing from an ordinary employee to a senior executive. Combining the influence weights of each transition state on the model's risk prediction results, a state transition probability matrix with each row summing to 1 is generated through normalization processing to ensure that the Markov evolution process conforms to the feature change patterns in real business scenarios.

[0024] This step employs a differential channel processing mechanism for heterogeneous features, effectively avoiding the backpropagation link breakage phenomenon caused by the non-differentiability of the input space containing discrete variables when conventional techniques are used. By separating the processing and implementing directional evolution in the target feature space, it is ensured that the algorithm execution logic can form a specific technical association with the inherent characteristics of the heterogeneous physical data.

[0025] In some embodiments, the continuous perturbation component and the discrete perturbation component are spliced ​​and fused, and filtered according to preset physical business logic constraints to generate a corresponding adversarial sample set, including: executing splicing instructions on the two components in the tensor splicing dimension; calling a preset rule engine library to compare the spliced ​​tensor elements with the legal value range boundary; wherein, the value range boundary interval of the feature field corresponding to the continuous perturbation component is extracted, and it is determined whether the continuous perturbation component exceeds the value range boundary interval.

[0026] Specifically, when the value exceeds the boundary interval of the value range, the continuous disturbance component is forcibly mapped to the extreme point of the boundary interval of the value range using a piecewise truncation function. The extreme point represents the highest or lowest limit value that the physical feature can exist in the real business environment. For example, the debt amount cannot be negative, and the minimum value after truncation is 0. For discrete disturbance components, it is verified whether the target state conforms to the business logic. For example, the income feature corresponding to the student occupation cannot exceed the preset upper limit. If it does not conform, the legal state with the second highest transition probability is selected as the replacement. Specifically, the preset physical business logic constraints are stored according to feature type. Continuous feature constraints include numerical range, dimensional consistency, and cross-feature logical association. For example, interest rate features must be within a fixed range, and monthly repayment amount cannot exceed a fixed percentage of monthly income. Discrete feature constraints include state legality and state combination rationality. For example, occupation type must be within a preset enumeration list, and no fixed income and high mortgage cannot coexist. The rule engine verifies and filters the perturbed features one by one, eliminating samples with logical distortion.

[0027] This filtering and truncation step is not a simple format check as in conventional techniques. Instead, it forces the purely mathematical vectors generated by the algorithm back into the real physical world, ensuring that the generated adversarial sample has the potential to actually launch an attack and ensuring that the sample construction and the business physical meaning are functionally mutually supportive.

[0028] In some embodiments, the basic risk control feature set and the adversarial example set are respectively input into the target risk control model to obtain the difference in predicted probability features between the two, and a confidence decay surface along the perturbation direction is constructed based on the difference in predicted probability features, including: performing multiple rounds of feature infusion and probability extraction actions; extracting the probability difference sequence and establishing a decay mapping relationship with the perturbation amplitude as the horizontal axis; wherein, the perturbation amplitude in the adversarial example set is subjected to multi-level discretization processing to generate multiple asymptotic perturbation scale vectors, and the asymptotic perturbation scale vectors are sequentially injected into the basic risk control feature set.

[0029] Specifically, the perturbation amplitude is used as the independent variable, the output probability vector is used as the dependent variable, and a polynomial of fixed order is used to perform fitting operations to generate the confidence decay surface. During the fitting process, the polynomial coefficients are optimized by the mean square error minimization criterion to ensure that the surface can accurately depict the continuous trend of the model confidence changing with the perturbation amplitude. Specifically, the effectiveness of the fitted confidence decay surface is verified by selecting the perturbation amplitude values ​​that were not involved in the fitting as test points, inputting them into the model to obtain the actual output probability, calculating the error between the actual value and the surface prediction value, and adjusting the polynomial order and refitting if the error exceeds a fixed proportion until the error meets the preset accuracy requirements, so as to ensure that the surface can truly reflect the model's anti-perturbation characteristics.

[0030] This step breaks away from the traditional approach of relying solely on static discrete points to calculate macroscopic indicators. By constructing a continuous confidence decay surface, it achieves dynamic topological mapping of the high-dimensional hidden decision boundaries of the model, providing objective data support for revealing underlying structural vulnerabilities.

[0031] In some embodiments, a critical perturbation threshold is extracted based on the confidence decay surface, and the critical perturbation threshold is converted into a dynamic regularization penalty term and injected into the loss function of the target risk control model. The network weights of the target risk control model are updated through backpropagation. This includes: analyzing the curvature feature points of the confidence decay surface to identify vulnerable locations; reconstructing the gradient guidance formula for backpropagation of the model; and updating the neuron connection strength along the reverse direction of the reconstructed gradient. Specifically, the second derivative feature of the confidence decay surface at the critical perturbation threshold is calculated, and the weighted product result is extracted to construct the dynamic regularization penalty term.

[0032] Specifically, the gradient mask matrix and the gradient tensor of the reconstruction loss function are subjected to Hadamard product operation to generate a corrected gradient tensor. The network weights are updated in the opposite direction of the corrected gradient tensor. An adaptive learning rate is used during the update process. The initial learning rate is set to a fixed value and decays with the number of iterations to avoid the model from overfitting due to excessive weight update. Specifically, the dynamic regularization penalty term is constructed based on the curvature characteristics of the confidence decay surface and the critical perturbation threshold, and the core calculation formula is as follows:

[0033] in, The target loss function after reconstruction, The original cross-entropy loss of the model, These are weighting coefficients that are dynamically adjusted according to the critical disturbance threshold. The second derivative of the confidence decay surface at the critical perturbation threshold represents the surface curvature. The L2 norm of the model network weights is used; by attaching this dynamic regularization penalty term to the original loss function, the constraints on the vulnerable regions of the model can be strengthened through curvature features, thus achieving targeted reinforcement. Specifically, the gradient mask matrix and the gradient tensor of the reconstruction loss function are subjected to Hadamard product operation to generate a corrected gradient tensor. The network weights are updated in the opposite direction of the corrected gradient tensor. An adaptive learning rate is used during the update process. The initial learning rate is set to a fixed value and decays with the number of iterations to avoid the model from overfitting due to excessive weight update.

[0034] Specifically, the distance sensitivity threshold is dynamically set according to the importance of the feature. For core risk control features such as user income and transaction frequency, the sensitivity threshold is set to a fixed value, and the cosine distance is less than this value and is judged as a non-sensitive channel. For secondary features, such as the account opening region, the sensitivity threshold is set to another fixed value to ensure that the gradient update of the core features is more targeted, while reducing the redundant calculation overhead of secondary features.

[0035] The introduction of dynamic regularization penalty terms and gradient mask filtering mechanism in this step completely changes the loss function construction rules and gradient backpropagation path of the underlying neural network. This design is not a uniform parameter decay as is the conventional technique in this field, but a customized surface smoothing operation is implemented for the vulnerable channels of the model. It effectively resolves the inherent coupling contradiction between high risk control detection rate and high anti-interference ability, and significantly improves the robustness and computational determinism of the computer system when performing high-dimensional feature discrimination tasks.

[0036] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An automated evaluation and adversarial example generation method for AI risk control models, characterized in that, include: Obtain the target risk control model and the basic risk control feature set, wherein the basic risk control feature set includes continuous feature dimension vectors and discrete feature dimension vectors; Jacobian gradient information is extracted from the parameter matrix of the target risk control model. Based on the Jacobian gradient information, gradient optimization is performed on the continuous feature dimension vector to generate continuous perturbation components. Based on the pre-constructed state transition probability matrix, Markov evolution is performed on the discrete feature dimension vector to generate discrete perturbation components. The continuous perturbation component and the discrete perturbation component are spliced ​​and fused, and filtered according to preset physical business logic constraints to generate a corresponding set of adversarial samples. The basic risk control feature set and the adversarial sample set are respectively input into the target risk control model to obtain the difference in their predicted probability features. Based on the difference in their predicted probability features, a confidence decay surface along the disturbance direction is constructed. The critical perturbation threshold is extracted based on the confidence decay surface, and the critical perturbation threshold is transformed into a dynamic regularization penalty term and injected into the loss function of the target risk control model. The network weights of the target risk control model are updated through backpropagation to complete the automatic reinforcement of the target risk control model.

2. The automated evaluation and adversarial example generation method for AI risk control models according to claim 1, characterized in that, Based on the Jacobian gradient information, gradient optimization is performed on the continuous feature dimension vector to generate continuous perturbation components, including: Extract the loss gradient vector of the target risk control model for the target risk category; Calculate the partial derivative matrix of the loss gradient vector in the feature space corresponding to the continuous feature dimension vector; The gradient ascent direction is determined based on the partial derivative matrix, and numerical accumulation is performed in the gradient ascent direction according to a preset iteration step size to generate continuous perturbation components.

3. The automated evaluation and adversarial example generation method for AI risk control models according to claim 1, characterized in that, Based on the pre-constructed state transition probability matrix, Markov evolution is performed on the discrete feature dimension vector to generate discrete perturbation components, including: Read the initial state category node of the discrete feature dimension vector; In the state transition probability matrix, query the transition probability value of each candidate state node corresponding to the initial state category node; The candidate state node with the highest transition probability value and which causes the risk score of the target risk control model to decrease by more than a preset decrease threshold is extracted as the target state, and the discrete disturbance component is generated based on the target state.

4. The automated evaluation and adversarial example generation method for AI risk control models according to claim 1, characterized in that, Filtering is performed based on preset physical business logic constraints, including: Extract the value range boundary interval of the feature field corresponding to the continuous disturbance component; Determine whether the continuous disturbance component exceeds the value range boundary interval; In cases where the value exceeds the boundary interval of the range, the continuous disturbance component is forcibly mapped to the extreme point of the boundary interval of the range using a piecewise truncation function.

5. The automated evaluation and adversarial example generation method for AI risk control models according to claim 1, characterized in that, Based on the predicted probability feature difference, a confidence decay surface is constructed along the perturbation direction, including: The perturbation amplitude in the adversarial sample set is discretized at multiple levels to generate multiple asymptotic perturbation scale vectors. The multiple progressive perturbation scale vectors are sequentially injected into the basic risk control feature set to obtain test sample matrices with different step dimensions; Record the output probability vector of the target risk control model when processing test sample matrices of various step dimensions, use the disturbance amplitude as the independent variable and the output probability vector as the dependent variable to perform polynomial fitting, and generate the confidence decay surface.

6. The automated evaluation and adversarial example generation method for AI risk control models according to claim 1, characterized in that, The critical perturbation threshold is transformed into a dynamic regularization penalty term and injected into the loss function of the target risk control model, including: Calculate the second derivative feature of the confidence decay surface at the critical perturbation threshold, where the second derivative feature characterizes the surface curvature; The surface curvature is linearly weighted with the original cross-entropy loss calculation term, and the weighted product result is extracted to construct the dynamic regularization penalty term. The dynamic regularization penalty term is appended to the end of the original loss function of the target risk control model to form a reconstructed loss function.

7. The automated evaluation and adversarial example generation method for AI risk control models according to claim 6, characterized in that, Before updating the network weights of the target risk control model via backpropagation, the following steps are also included: Extract the forward propagation intermediate tensor representation of the hidden layers within the target risk control model; The forward propagation intermediate tensor is calculated as the cosine distance vector between the input basic risk control feature set and the input adversarial sample set; The elements in the cosine distance vector that are less than the distance-sensitive threshold are indexed and marked as non-sensitive channels.

8. The automated evaluation and adversarial example generation method for AI risk control models according to claim 7, characterized in that, Updating the network weights of the target risk control model through backpropagation includes: A binary gradient mask matrix is ​​generated based on the non-sensitive channel; When calculating the gradient tensor of the reconstruction loss function relative to the hidden layer weight matrix, the gradient tensor is multiplied by the gradient mask matrix to generate a modified gradient tensor. The network weights are updated in the opposite direction of the corrected gradient tensor.

9. The automated evaluation and adversarial example generation method for AI risk control models according to claim 1, characterized in that, After inputting the basic risk control feature set and the adversarial example set into the target risk control model, the method further includes: The basic accuracy features output by the target risk control model based on the basic risk control feature set and the adversarial accuracy features output by the adversarial sample set are obtained respectively. Calculate the ratio of the base accuracy feature to the adversarial accuracy feature, and use the ratio as the adversarial sensitivity weight; The initial Kolmogorov-Smirnov statistic, calculated based on the empirical cumulative distribution function, is feature-scaled using the adversarial sensitivity weights to generate automated evaluation indicators.

10. The automated evaluation and adversarial example generation method for AI risk control models according to claim 1, characterized in that, The process of updating the network weights of the target risk control model through backpropagation also includes: The norm of the weight difference between two adjacent updates is calculated according to a preset cycle. If the weight difference norm is less than the weight convergence threshold within multiple consecutive cycles, the backpropagation update link is truncated, and the current network weight is frozen.