Data Access Control and Authentication Methods for Intelligent Systems Based on Deep Learning

Abstract

This invention discloses a data access control and authentication method for intelligent systems based on deep learning. This invention relates to the field of identity authentication and access control technology. The method collects behavioral characteristics of the accessing subject and inputs them into a causal deep authentication network to obtain a security confidence score. It then calculates the average causal effect through feature node intervention to generate a feature attribution map, enabling interpretable authentication decisions. Semantic encoding of the task description text yields an intent vector, which is matched with a permission policy tree to determine the data permission range. A micro-isolation sandbox token with a preset lifespan is generated, enabling dynamic fine-grained authorization. The confidence score is divided based on high and low confidence thresholds. Suspected drifting samples undergo secondary verification and are consolidated with elastic weights for incremental fine-tuning, balancing behavioral adaptation and anti-forgetting. This invention is applicable to data access security protection under a zero-trust architecture.

Description

Technical Field

[0001] This invention relates to the field of identity authentication and access control technology, specifically to a data access control and authentication method for intelligent systems based on deep learning. Background Technology

[0002] With the rapid evolution of intelligent systems such as high-density networks, the Industrial Internet, and large-scale artificial intelligence models, data access control and identity authentication have become core aspects of ensuring system security. Role-based access control and attribute-based access control are both static policy-driven models, with authorization granularity typically based on data tables or system modules. This makes them unsuitable for the fine-grained permission requirements of high-concurrency dynamic interactions between machines. Zero-trust architecture further introduces the concept of continuous verification, combining deep learning models to dynamically assess the trust of access subjects' behavioral characteristics. The closest existing technology to this invention is a deep learning-based continuous authentication method based on user behavior. This method collects behavioral log features of access subjects to construct feature vectors, which are then input... The pre-trained deep neural network end-to-end fits to obtain a trust score. The trust score is compared with a preset fixed threshold to make a decision to allow or deny, and the decision result is used as a supplementary control signal in addition to the static authorization strategy. The above-mentioned existing technology only relies on the end-to-end fitting of trust scores by the neural network for binary classification. It does not model the causal relationship between input features and output results, nor does it dynamically shrink the scope of permissions in combination with real-time task semantics. Moreover, it adopts a static training mode with fixed parameters. Therefore, there are problems such as uninterpretable authentication decisions, coarse authorization granularity, and difficulty in adapting the model to changes in behavior. In addition, in multi-agent cross-domain collaborative calling scenarios, the above problems will further exacerbate the risks of data leakage and compliance audit. Summary of the Invention

[0003] The technical problem this invention aims to solve is that existing deep learning-based intelligent system continuous authentication methods rely solely on trust scores obtained through end-to-end fitting of neural networks for binary classification, failing to explicitly model the causal relationship between input features and output results. This leads to uninterpretable authentication decisions, making it difficult to meet the interpretability requirements of security audits. Existing access control mechanisms mostly adopt static policy configuration modes based on roles or attributes, without combining real-time task semantics to dynamically shrink the scope of permissions in a fine-grained manner, which easily leads to over-authorization and unauthorized access risks. Existing deep learning authentication models use static training modes with fixed parameters, failing to consider the reality of legitimate user behavior evolving over time. If online full retraining is adopted for adaptation, it is prone to catastrophic forgetting, creating a contradiction between adapting to new behaviors and retaining historical defense capabilities.

[0004] To address the aforementioned technical issues, this invention provides a data access control and authentication method for intelligent systems based on deep learning. This method first collects the behavioral characteristics of the accessing entity, constructs a structured feature vector containing operation frequency, geographical location offset, access time period, requested resource type, and historical behavior similarity, and inputs it into a pre-trained causal deep authentication network. This network consists of a feature encoding layer and a structured causal graph layer. While outputting a security confidence score, it performs intervention operations on each feature node in the feature vector, fixing its value to an abnormal state and fixing its value to a normal state, obtaining corresponding counterfactual outputs. Then, it calculates the average causal effect of each feature node and generates an auditable and traceable feature attribution graph, sorted by contribution, thereby achieving interpretable output for authentication decisions.

[0005] Furthermore, this invention semantically encodes the contextual task description text of the access request to obtain an intent vector, calculates the similarity between the intent vector and the intent vectors of each policy node in the pre-built permission policy tree, selects the policy node with the highest matching degree to determine the data permission range that matches the current task intent, and generates a micro-isolation sandbox token with a preset life cycle accordingly. The token is forcibly revoked after the task is completed or the life cycle expires, thereby realizing dynamic fine-grained authorization with permission range accurate to the data row and column level.

[0006] Furthermore, this invention sets two confidence thresholds, high and low, and divides the security confidence score into three intervals: normal access, suspected drift, and high-risk rejection. Samples in the suspected drift interval trigger permission downgrade and multimodal secondary verification. Samples that pass and fail the verification are respectively assigned to the incremental training sample buffer pool and the historical malicious sample library. Based on the elastic weight consolidation mechanism, an incremental fine-tuning total loss function is constructed to update the parameters of the causal deep authentication network, so as to adapt to the latest legitimate behavior characteristics of the access subject while retaining the ability to identify historical malicious characteristics.

[0007] The confidence score output by the aforementioned causal interpretable authentication module, together with the feature attribution result, drives the permission contraction decision of the dynamic micro-segmentation weight reduction module. After the dual-threshold adaptive incremental learning module performs anti-forgetting fine-tuning based on the judgment result of suspected drift samples in this closed loop, its updated parameters are fed back to the causal interpretable authentication module to form a closed loop. This cross-module causal result reverse-drives the collaborative mechanism of permission contraction and incremental learning, which overcomes the technical bias that trust scoring, permission policy and model training are independent and do not interact with each other in the existing technology, and achieves an overall security gain that cannot be achieved by implementing the three individually.

[0008] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0009] Firstly, by using counterfactual intervention calculation and feature attribution mapping, the authentication system can simultaneously provide a structured, auditable, and traceable explanation of the reasons while outputting a rejection decision, thus meeting the security audit requirements in scenarios with high compliance requirements.

[0010] Secondly, by matching the similarity between intent vectors and policy trees, micro-isolation sandbox tokens with extremely short lifecycles and permission ranges accurate to the data row and column level are dynamically generated. After the task is completed, they are forcibly reclaimed in a timely manner. By precisely limiting the token lifecycle and permission range, the attack surface is narrowed, reducing the risk of unauthorized access and data leakage.

[0011] Third, through an incremental learning mechanism that combines dual threshold judgment with elastic weight consolidation, the model adapts to the latest legitimate behavior features while retaining the ability to identify historically known malicious attack features, thus enabling the authentication model to continuously maintain its ability to identify historical malicious features. Attached Figure Description

[0012] Figure 1 This is a closed-loop schematic diagram of the overall system architecture of the present invention;

[0013] Figure 2 This is a flowchart of the causal deep learning authentication module of the present invention;

[0014] Figure 3 This is a schematic diagram of an embodiment of the closed-loop process of the overall method of the present invention. Detailed Implementation

[0015] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0016] The data access control and authentication method for intelligent systems based on deep learning provided by this invention consists of a behavior feature acquisition layer, a causal deep learning authentication module, a dynamic micro-segmentation weight reduction module, a dual-threshold adaptive anti-forgetting incremental learning module, a policy and vector storage layer, a token management and caching layer, and an audit log output layer, forming a closed-loop system. The behavior feature acquisition layer is responsible for collecting raw data such as operation logs, geographical location, access frequency, and contextual task descriptions of the access subject (user or AI agent), and sending them to the causal deep learning authentication module and the dynamic micro-segmentation weight reduction module respectively. The security confidence score output by the causal deep learning authentication module is sent to the dual-threshold adaptive anti-forgetting incremental learning module for judgment and diversion. Access requests that are normally allowed and those that pass the second verification enter the dynamic micro-segmentation weight reduction module to generate micro-segmentation sandbox tokens. Samples suspected of drifting and high-risk rejection branches flow into the sample buffer pool and malicious sample library of the dual-threshold adaptive anti-forgetting incremental learning module respectively for offline incremental fine-tuning. The feature attribution map and all judgment results are finally merged into the audit log output layer to form a complete closed loop.

[0017] I. Causal Deep Learning Authentication Module

[0018] This module first collects the original behavioral characteristics of the current user, constructs a structured feature vector, denoted as... The feature dimensions include operation frequency, geographic location offset, access time period, requested resource type, and historical behavior similarity. The feature vector is then used to... The input is a pre-trained causal deep authentication network, which consists of a Transformer feature encoding layer and a structural causal graph layer. The structural causal graph layer pre-learns a directed acyclic graph representing the causal dependencies between feature nodes based on domain knowledge and historical data. The network outputs an initial security confidence score. And the causal characteristics of the intermediate layer.

[0019] For samples whose judgment results fall into the rejection interval or suspected drift interval, for each feature node Perform intervention operations with values ​​fixed as abnormal states respectively. Intervention operations with values ​​fixed to the normal state This yields the corresponding counterfactual output. The causal quantification relationship upon which this module is based is:

[0020] ;

[0021] in, Representation of features Average causal effect This represents the access decision result output by the authentication module. The allowed value is 1, and the denied value is 0. Represents the characteristic variable Intervention was applied to fix it to an abnormal state value. Represents the characteristic variable Intervention is applied to fix it to its normal value. This indicates the expectation of the model's output.

[0022] The counterfactual output is calculated as follows:

[0023] ;

[0024] in, Indicates features Replace with counterfactual values The output obtained by recalculating the model later This represents the overall mapping function of the trained causal deep authentication network. Represents the components of the original input feature vector. Representation of features The corresponding counterfactual value.

[0025] according to Sort all feature nodes in descending order by absolute value, select the top-ranking feature nodes by contribution, and calculate the normalized contribution percentage of each feature:

[0026] ;

[0027] in, Representation of features The percentage of attributable contributions to the current refusal to make a decision. This represents the total number of features participating in the causal attribution calculation. The denominator is the sum of the absolute values ​​of the average causal effects of all features, used to normalize the contribution of each feature. Features with the highest contribution ranking and their proportions are mapped to a preset structured natural language log template, generating a feature attribution map that can be traced by security audits. This map, along with the judgment result, is output to the audit log layer for storage. In one specific embodiment, the total number of features... The value is set to 20, and the number of features with the highest contribution ranking is set to 5. The generated log template records the name of each feature, the corresponding contribution percentage, and the final access judgment result, so that compliance auditors can trace the reasons for rejection.

[0028] II. Dynamic Micro-Issuance Weighting Module

[0029] This module obtains the contextual task description text of the current access request, including the interface call parameters initiated by the AI ​​agent, task instructions, and call chain context information. It then uses a lightweight pre-trained language model fine-tuned with domain data to semantically encode the text, resulting in an intent vector. , will the intention vector Input vector database, and intent vectors of each policy node in the pre-built permission policy tree. Calculate the cosine similarity one by one:

[0030] ;

[0031] in, This represents the intent vector obtained by semantic encoding of the current access request. This represents the intent vector corresponding to the predefined policy node in the permission policy tree. Represents the dot product of two vectors. and Let represent the Euclidean norms of the two vectors, respectively. The value ranges from negative one to positive one. The larger the value, the higher the degree of matching between the current task intent and the strategy node.

[0032] The nodes are sorted from highest to lowest similarity scores. The first node with a similarity score exceeding a preset matching threshold is selected as the target policy node. The data permission scope corresponding to this node is determined, which clearly defines the allowed data tables, data row ranges, data column fields, and executable operation types. Combining the security confidence score output by the causal deep learning authentication module with the judgment result, if the overall judgment is that access is allowed, a micro-isolation sandbox token is dynamically generated based on the above data permission scope. An extremely short lifespan is adaptively set according to the estimated task execution time. The token encapsulates the permission scope, lifespan, and subject identification information, is issued to the access subject in JWT format, and is written to a distributed cache for unified management. When the task is completed or the lifespan expires, the system forcibly reclaims and cancels the token, and clears the corresponding permission record in the cache to ensure that the permission expires immediately. In a specific embodiment, the matching threshold is set to 0.75, and the lifespan is adaptively set within the range of 30 seconds to 600 seconds based on the estimated task execution time.

[0033] III. Dual-threshold adaptive anti-forgetting incremental learning module

[0034] This module continuously monitors the security confidence scores of each accessing entity during the online authentication process. Based on the high confidence threshold With low confidence threshold Make a judgment: when When, it is determined that access is allowed. When this is detected as a suspected drift, permission downgrade and multimodal secondary verification are triggered. When access is denied, the access is determined to be downgraded. This means that the data permissions of the accessing subject are tightened to the read-only permission range. The secondary verification includes biometric recognition, dynamic password verification, and behavior secondary confirmation. Samples that pass the secondary verification are marked as legitimate drift feature samples and added to the incremental training sample buffer pool. Samples that fail the secondary verification are marked as suspected attack feature samples and added to the historical malicious sample library. At the same time, the access is directly denied.

[0035] During periods of low system load, the optimal parameter values ​​are stored and fixed after training on historical tasks. With the corresponding Fisher information matrix weights By combining the newly added legitimate drift samples in the buffer pool, an incremental fine-tuning total loss function is constructed:

[0036] ;

[0037] in, This represents the total loss function during the incremental fine-tuning phase. This represents the cross-entropy loss calculated by the model on newly acquired legal drift samples. This represents the vector of model parameters to be optimized. For its first One portion, This indicates the first historical task training completed and the data is now permanently saved. Optimal values ​​for each parameter This indicates the importance weight of the parameter in the Fisher information matrix. The penalty intensity hyperparameter is used to balance the trade-off between learning new knowledge and retaining historical knowledge.

[0038] The backpropagation algorithm is used to analyze the total loss function. Gradient descent optimization is performed, applying stronger offset constraints to parameter components with higher Fisher information weights. This allows the model to adapt to the latest legitimate behavior features while retaining its ability to identify historically known malicious attack features. After fine-tuning, the new model parameters and the updated Fisher information matrix are re-established and saved as the benchmark for the next round of incremental learning. The dual thresholds are then dynamically adjusted based on recent statistical feedback from false rejection and false acceptance rates. , The value of is used to complete a full continuous learning loop. In a specific embodiment, the initial value of the high confidence threshold is 0.85, the initial value of the low confidence threshold is 0.40, and the penalty intensity hyperparameter is... The value is 100.

[0039] IV. Overall Method Closed-Loop Process Implementation Example

[0040] Taking a multi-agent cross-domain collaborative invocation scenario in the core production environment of a financial institution as an example: An AI agent initiates a query request for a customer transaction data table. The behavior feature collection layer collects features such as the agent's operation frequency, source network address, and access time period to construct a feature vector. The feature vector is input into the causal deep authentication network and the security confidence score is 0.62, which is between the high and low confidence thresholds. It is judged as a suspected drift. The system then tightens the agent's permissions to the read-only scope and triggers a secondary multimodal verification. After the verification passes, the dynamic micro-segmentation demotion module performs semantic encoding on the task instruction text submitted by the agent to obtain an intent vector. The similarity is calculated with each policy node in the permission policy tree to determine the similarity with the current task. The data access scope of the matched target strategy node is limited to the transaction record row range and specified field column corresponding to the specific customer number. A micro-isolation sandbox token with a lifespan of 120 seconds is generated and issued in JWT format. After the agent completes the query task with the token, the token is forcibly reclaimed. The samples identified as suspected drifts are marked and stored in the incremental training sample buffer pool for the system to perform offline incremental fine-tuning in combination with the Fisher information matrix weights during low load periods, and the dual threshold values ​​are adjusted accordingly for the judgment of subsequent access requests. At the same time, the feature attribution map generated by the causal deep learning authentication module during this access process is written to the audit log layer along with all judgment results for compliance auditors to trace and review.

[0041] It should be noted that, in addition to the aforementioned interventional calculation method based on structural causal graphs, the causal attribution calculation can also employ a calculation method based on the marginal contribution mean of feature subsets to achieve equivalent interpretability output in other embodiments. Similarly, in addition to using a lightweight pre-trained language model, the deep intent understanding module can also employ an intent matching method based on knowledge graphs and rule reasoning engines in other embodiments. This method determines the data permission range by performing structured parsing of access request parameters and matching them with a pre-built rule mapping table. This approach is more suitable for scenarios with relatively fixed task types and clear business rules. All of the above alternative implementation methods can achieve the technical effects of this invention and fall within the scope of protection of this invention.

Claims

1. A data access control and authentication method for intelligent systems based on deep learning, characterized in that, include: Collect the behavioral characteristics of the accessing entities and construct a structured feature vector. The structured feature vector includes operation frequency, geographical location offset, access time period, requested resource type, and historical behavior similarity. The structured feature vector is input into a pre-trained causal deep authentication network to obtain a security confidence score and an intermediate causal feature representation. The causal deep authentication network includes a feature encoding layer and a structural causal graph layer. The structural causal graph layer learns a directed acyclic graph based on domain knowledge and historical data to represent the causal dependency connections between feature nodes. For each feature node in the structured feature vector corresponding to an access request whose judgment result is access denied or suspected drift, an intervention operation with a fixed value of abnormal state and an intervention operation with a fixed value of normal state are executed respectively to obtain the counterfactual outputs corresponding to the two intervention operations. The average causal effect value of the feature node is calculated based on the difference between the two counterfactual outputs. All feature nodes are sorted according to the absolute value of the average causal effect value. The top-ranked feature nodes are selected, and the normalized contribution ratio of each of the feature nodes is calculated. The normalized contribution ratio is mapped to a preset structured natural language log template to generate a feature attribution map. The context task description text of the access request is obtained, and the context task description text is semantically encoded using a natural language processing model to obtain an intent vector. The similarity between the intent vector and the intent vector of each policy node in the pre-constructed permission policy tree is calculated. The similarity values ​​are sorted from high to low, and the policy node with the first ranking and the similarity value exceeding the preset matching threshold is selected as the target policy node. The data permission range corresponding to the target policy node is determined, and the data permission range limits the data tables, data row ranges, and data column fields that are allowed to be accessed. The security confidence score is determined based on a high confidence threshold and a low confidence threshold. Access requests with a security confidence score greater than or equal to the high confidence threshold are allowed access, access requests with a security confidence score less than the low confidence threshold are denied access, and access requests with a security confidence score between the low confidence threshold and the high confidence threshold are suspected of drifting. Access requests suspected of drifting are subject to permission downgrading and secondary verification. For access requests that are determined to be allowed and access requests that pass the secondary verification, a micro-isolation sandbox token with a preset lifespan is generated according to the data permission scope, and the micro-isolation sandbox token is issued to the accessing subject. When either the task is completed or the lifespan expires, the micro-isolation sandbox token is revoked. Samples corresponding to access requests that pass the second verification are stored in the incremental training sample buffer pool, samples corresponding to access requests that fail the second verification are stored in the historical malicious sample library, and the corresponding access requests are directly rejected. Based on the samples in the incremental training sample buffer pool, the parameter values ​​that have been fixed and saved after the training of historical tasks, and the corresponding Fisher information matrix weights, an incremental fine-tuning total loss function is constructed. The incremental fine-tuning total loss function is optimized by gradient descent to update the parameters of the causal deep authentication network. In this process, a stronger offset constraint is applied to the parameter components with larger Fisher information matrix weight values.

2. The method according to claim 1, characterized in that, The abnormal state value and the normal state value are determined based on the statistical distribution of historical samples corresponding to the feature node.

3. The method according to claim 1, characterized in that, The feature attribution map records the feature name, the normalized contribution ratio, and the access determination result, and outputs them synchronously to the audit log layer for retention.

4. The method according to claim 1, characterized in that, The natural language processing model is a pre-trained language model fine-tuned with domain data, and the context task description text includes interface call parameters initiated by the accessing subject, task instructions, and call chain context information.

5. The method according to claim 1, characterized in that, The micro-segmentation sandbox token encapsulates the data permission scope, the lifespan, and the access subject identification information. It is issued in JWT format and written into a distributed cache for management. The lifespan is set according to the estimated execution time of the task.

6. The method according to claim 1, characterized in that, The permission downgrade refers to tightening the data permissions of the accessing subject to the read-only permission range, and the secondary verification includes biometric identification, dynamic password verification, and secondary confirmation of behavior.

7. The method according to claim 1, characterized in that, Also includes: After updating the parameters of the causal deep authentication network, the updated parameter values ​​and the updated Fisher information matrix weights are solidified and saved as a benchmark for subsequent incremental learning. Based on the statistical feedback of recent false rejection and false acceptance rates, the values ​​of the high confidence threshold and the low confidence threshold are adjusted.

8. The method according to claim 1, characterized in that, The total loss function for incremental fine-tuning is: ； in, This represents the incremental fine-tuning total loss function. This represents the cross-entropy loss calculated on the samples in the incremental training sample buffer. This represents the parameter vector to be optimized in the causal deep authentication network. For its first One portion, This indicates the first historical task training completed and the data is now permanently saved. Each parameter value This indicates the weight of the parameter component in the Fisher information matrix. This is the penalty intensity hyperparameter.

9. A data access control and authentication system for intelligent systems based on deep learning, characterized in that, include: The behavior feature acquisition module is used to collect the behavior features of the accessing entities and construct structured feature vectors. The causal deep learning authentication module is used to input the structured feature vector into a pre-trained causal deep authentication network to obtain a security confidence score. For each feature node in the structured feature vector corresponding to an access request whose judgment result is denied access or suspected drift, an intervention operation with a fixed value of abnormal state and an intervention operation with a fixed value of normal state are performed respectively. The average causal effect value of the feature node is calculated based on the obtained counterfactual output, and then a feature attribution map is generated. The dynamic micro-segmentation demotion module is used to semantically encode the context task description text of the access request to obtain an intent vector, calculate the similarity between the intent vector and the intent vector of each policy node in the permission policy tree to determine the data permission range, and generate a micro-segmentation sandbox token with a preset life cycle based on the data permission range. The dual-threshold adaptive anti-forgetting incremental learning module is used to determine the safety confidence score based on high confidence threshold and low confidence threshold, and to incrementally fine-tune the parameters of the causal deep authentication network based on samples in the suspected drift interval. The output of the causal deep learning authentication module is connected to the dynamic micro-isolation weighting module and the dual-threshold adaptive anti-forgetting incremental learning module, respectively. The updated parameters of the dual-threshold adaptive anti-forgetting incremental learning module are fed back to the causal deep learning authentication module, forming a closed loop.

Images