Tobacco supervision platform access control method based on machine learning

Abstract

The invention discloses a tobacco supervision platform access control method based on machine learning, and the method comprises the following steps: collecting access user data, carrying out the alignment through a unified timestamp, and forming an access state data set; constructing a multi-level feature vector, and encoding the multi-level feature vector into a state vector; inputting the state vector into the improved neural circuit strategy model, and outputting an access control decision; carrying out conflict judgment on the access control decision, a static strategy library, a forced closing scale block and a time-space quota account book to form a synthetic access control result; executing an access operation on a gateway layer or a database agent layer of the tobacco supervision platform, and generating an audit log carrying evidence information; and carrying out on-line fine tuning or off-line retraining. According to the invention, the combination of the main decision and the fine-grained control is realized through the improved neural circuit strategy, and the safety, interpretability and adaptivity of the tobacco supervision platform are improved.

Description

Technical Field

[0001] This invention relates to the field of network and information security technology, and in particular to an access control method for a tobacco regulatory platform based on machine learning. Background Technology

[0002] With the continuous deepening of information technology construction in the tobacco industry, tobacco regulatory platforms play an important role in cross-departmental collaboration, data sharing, and risk prevention and control. Existing platforms mostly adopt mechanisms such as role-based access control (RBAC) and attribute-based access control (ABAC), using static rules to limit the access relationships between users and resources.

[0003] However, the above methods have obvious shortcomings: on the one hand, the number of access control policies is huge and difficult to update, and rule conflicts and coverage blind spots occur from time to time; on the other hand, traditional methods only make judgments based on user roles or static attributes, and cannot model the dynamic context of access behavior, making it difficult to detect risky behaviors such as batch export and access during abnormal periods in a timely manner.

[0004] Existing methods also have shortcomings in terms of fine-grained control and interpretability. They can usually only provide coarse-grained results of "allow or deny", which makes it difficult to support flexible management needs such as field desensitization, watermark embedding, and rate limiting. Although some black-box methods that introduce machine learning can improve detection accuracy, they lack traceability and adaptive evolution capabilities, making it difficult to meet the compliance and audit requirements of tobacco regulation.

[0005] Therefore, how to provide a machine learning-based access control method for tobacco regulatory platforms is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a machine learning-based access control method for tobacco regulatory platforms. This invention combines domain prior knowledge with machine learning by introducing a templated pre-wiring structure, learnable gating units, and a time-memory loop into a neural circuit policy model. It not only outputs the master decision of allowing, denying, or challenging, but also generates fine-grained control parameters such as field anonymization, watermark embedding, rate limiting, pagination constraints, and export limits. This constructs a closed-loop access control system encompassing data acquisition, state vector encoding, model inference, conflict resolution, auditing, and adaptive evolution, thereby significantly improving the security, interpretability, and maintainability of access control for tobacco regulatory platforms.

[0007] An access control method for a tobacco regulatory platform based on machine learning according to an embodiment of the present invention includes the following steps: Collect user access data, align it with a unified timestamp, and form an access status dataset; A multi-level feature vector is constructed based on the access state dataset, and the multi-level feature vector is encoded into a state vector; The state vector is input into the improved neural circuit strategy model, which consists of sensory neurons, inhibitory neurons, instruction neurons and interpreting neurons. It has a templated pre-wired structure, learnable gating units and time memory loops, and is used to output access control decisions. The access control decision is reconciled with the static policy library, the mandatory compliance module and the time-space quota ledger. If there is a conflict, the compliance and quota constraints are executed first to form a composite access control result. Based on the synthesized access control results, access operations are performed at the gateway layer or database proxy layer of the tobacco regulatory platform, and audit logs carrying evidence information are generated. Based on the audit logs and manual verification results, the improved neural circuit strategy model is fine-tuned online or retrained offline.

[0008] Furthermore, the formation of the access state dataset includes: The accessed user data includes user account attributes, organizational position information, historical operation logs, device fingerprints, network environment, geographical location, time characteristics, resource sensitivity annotations, and metadata of regulatory rules; Align data entries from different sources according to a globally unified timestamp; Perform quality checks on the collected data, removing duplicate logs, invalid fields, and obvious outliers; The cleaned data entries are organized into five dimensions: user, operation, resource, environment, and time; Using "user ID - access timestamp - operation ID" as a combined index, the data from the above five dimensions are concatenated into an access status record. All records are stored in chronological order to form an access status dataset.

[0009] Furthermore, the generation of the state vector includes: The multi-level feature vector includes user features, operation features, resource features, environmental features, and time-series features; User account attributes and organizational job information are extracted from the user dimension of the access status dataset to form user features; Operation characteristics are formed by extracting operation types, call frequencies, and concurrency from historical operation logs from an operational perspective. The sensitivity level, desensitization requirements, and cross-institutional attributes in resource sensitivity labeling are extracted from the resource dimension to form resource features; From an environmental perspective, device fingerprints are extracted along with trustworthiness, IP reputation score, and geographical location offset within the network environment to form environmental characteristics; Extract the time period, holiday markers, and nighttime access markers from the time dimension to form time features; The user features, operation features, resource features, environmental features, and time features are encoded and concatenated to generate the corresponding state vector.

[0010] Furthermore, the generation of the access control decision includes: The system receives the processing results of the state vector by the instruction neuron layer of the improved neural circuit strategy model and outputs the initial judgment signal. The initial determination signal is converted into an access control master decision via the main output unit, and the master decision is one of three options: allow, deny, or challenge. While generating the main decision, fine-grained control parameters are generated by multiple auxiliary output units of the instruction neuron layer. These fine-grained control parameters consist of one or more of the following five types of parameters: Field desensitization parameters are used to mask or obfuscate sensitive fields; Watermark embedding parameters are used to insert traceability identifiers during data export; Rate limiting parameters are used to limit the maximum frequency of access requests; Pagination constraint parameters are used to split large-scale queries into multiple pagination requests; The maximum number of records to export parameter is used to set the maximum number of records that can be exported in a batch. The master decision is combined with the fine-grained control parameters to form an access control decision vector, which is then output as the final access control decision result to guide access execution and auditing.

[0011] Furthermore, the improved neural circuit strategy model includes: The sensory neuron includes multiple input neurons, which correspond to user features, operation features, resource features, environmental features and time features, respectively, and are used to receive component inputs of the state vector; Suppressing neurons are configured with several suppression units corresponding to high-risk access conditions. The suppression units are connected to the perceptual neuron layer through a fixed pre-wiring method and are used to suppress the signal when a preset feature combination is detected. The instruction neuron includes at least three output units, which are used to generate allow signals, rejection signals and challenge signals, respectively, and are further connected to fine-grained control parameter output units to generate field desensitization, export restrictions, rate thresholds and watermark parameters; The explanatory neuron, which includes several neuronal units connected to inhibitory neurons and sensory neurons, is used to summarize the input contribution and trigger rule identifiers and output audit evidence information; Temporal memory loops establish circular connections between the perceptual neuron layer and the inhibitory neuron layer to maintain short-term contextual information of access behavior; Learnable gating units are inserted into partial connections between the perceptual neuron layer and the inhibitory neuron layer, and between the inhibitory neuron layer and the instruction neuron layer, to adjust the connection weight strength during the training phase.

[0012] Furthermore, the conflict resolution includes: The access control decision is compared with the preset access control rules in the static policy library. When the results are inconsistent, the mandatory constraint rules in the static policy library shall be given priority. The access control decision after comparison is checked against the regulatory red line rules in the compliance module. When a violation of mandatory compliance requirements occurs, the access control decision is adjusted to deny or challenge. Access control decisions are validated based on a spatiotemporal quota ledger. The spatiotemporal quota ledger records available access quotas using resource identifiers and time slices as indexes. When an access request is detected to exceed the remaining quota of the corresponding resource, the access control decision is downgraded to pagination execution, rate limiting execution, or direct rejection. After completing the comparison, verification, and quota check, the synthesized access control result with conflict resolution is output.

[0013] Furthermore, the generation of the audit log includes: When implementing access control decisions after conflict resolution, the user identifier, resource identifier, operation type, and execution time of the access request are recorded; The interpretation neuron layer of the improved neural circuit strategy model is invoked to extract the triggered access control rules, key feature contribution, and quota ledger status, which are then output as evidence information. The access request information is combined with the evidence information to generate a complete audit log entry; The audit log entries are stored in the audit log database.

[0014] Furthermore, the online fine-tuning and offline retraining include: Based on the audit logs and manual verification results, the execution feedback of access control decisions is extracted, and cases of accidental access and accidental blocking are marked. During the online operation phase, the learnable gating units of the improved neural circuit strategy model are fine-tuned using a small-step parameter update method to adapt to the latest access behavior patterns. During the offline training phase, newly added audit log samples and feedback labels are included in the training set, and the improved neural circuit strategy model is retrained in combination with historical data to update connection weights and rule mappings. After the improved neural circuit strategy model is updated, a verification is performed to ensure that the output of the improved neural circuit strategy model does not violate the mandatory constraints of the compliance module, and the online model is gradually replaced through a canary release method.

[0015] The beneficial effects of this invention are: This invention achieves an organic combination of domain prior rules and machine learning models by introducing templated pre-wiring structures, learnable gating units, and temporal memory loops into the neural circuit policy model. This ensures compliance and interpretability while dynamically adapting to new access behavior patterns, thus enhancing the model's robustness and practicality.

[0016] The access control method proposed in this invention can not only output the master decision of "allow, deny, or challenge", but also generate fine-grained control parameters such as field desensitization, watermark embedding, rate limiting, pagination constraints, and export limits, thereby realizing a multi-layered security protection mechanism and avoiding the problem of excessively coarse granularity in traditional access control.

[0017] This invention constructs a closed-loop system that includes access state dataset generation, state vector encoding, neural circuit policy reasoning, conflict resolution, audit evidence output, and adaptive evolution. It can achieve dynamic optimization of access control policies while ensuring compliance, and has higher security, interpretability, and maintainability. It is suitable for complex application scenarios of tobacco regulatory platforms. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine learning-based access control method for a tobacco regulatory platform proposed in this invention. Figure 2 This is a diagram of an improved neural circuit strategy model for an access control method for a tobacco regulatory platform based on machine learning, as proposed in this invention. Figure 3 This is a data flow diagram of an access control method for a tobacco regulatory platform based on machine learning proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1-3 A machine learning-based access control method for a tobacco regulatory platform includes the following steps: Collect user access data, align it with a unified timestamp, and form an access status dataset; A multi-level feature vector is constructed based on the access state dataset, and the multi-level feature vector is encoded into a state vector; The state vector is input into the improved neural circuit strategy model, which consists of sensory neurons, inhibitory neurons, instruction neurons and interpreting neurons. It has a templated pre-wired structure, learnable gating units and time memory loops, and is used to output access control decisions. The access control decision is reconciled with the static policy library, the mandatory compliance module and the time-space quota ledger. If there is a conflict, the compliance and quota constraints are executed first to form a composite access control result. Based on the synthesized access control results, access operations are performed at the gateway layer or database proxy layer of the tobacco regulatory platform, and audit logs carrying evidence information are generated. Based on the audit logs and manual verification results, the improved neural circuit strategy model is fine-tuned online or retrained offline.

[0021] In this embodiment, the formation of the access status dataset includes: The accessed user data includes user account attributes, organizational position information, historical operation logs, device fingerprints, network environment, geographical location, time characteristics, resource sensitivity annotations, and metadata of regulatory rules; To ensure consistency of data from different sources, all collected data entries need to be aligned using a globally unified timestamp. This means setting a unified time base within the platform and uniformly marking user information, operation logs, device and network information, geographical location, and resource annotations in chronological order. This ensures that data from the same access request in different sources can be mapped to the same moment, thus establishing a correlation.

[0022] After alignment, the collected data undergoes quality verification, including checking for duplicate logs, invalid fields, and obvious outliers. Duplicate logs refer to redundant records with completely identical content; only one such record is retained in the dataset. Invalid fields refer to entries with missing values ​​or incorrect formats, and are removed. Obvious outliers refer to records that do not conform to logic, such as reversed timestamps or geographical locations that deviate significantly from commonly used locations; such entries are not included in the normal dataset.

[0023] After cleaning, the entries are further organized according to five dimensions: user, operation, resource, environment, and time. In implementation, user account attributes and organizational position information are categorized into the user dimension to represent the accessing subject; historical operation logs are categorized into the operation dimension to reflect access behavior; resource sensitivity annotations are categorized into the resource dimension to describe the sensitivity level and attributes of the access target; device fingerprints and network environment are categorized into the environment dimension to reflect the terminal and network conditions; and geographical location and time characteristics are categorized into the time dimension to describe the spatiotemporal context of the access. Through this allocation rule, each access record is decomposed into five fixed dimensional structures, ensuring consistency in the organization of different access requests.

[0024] Using user ID, access timestamp, and operation ID as a combined index, information from these five dimensions is concatenated into an access state record and stored in chronological order. As access requests continue to arrive, records are continuously appended, eventually forming a complete access state dataset. This dataset provides a unified data foundation for the subsequent formation of state vectors and model inference.

[0025] In this embodiment, the generation of the state vector includes: The multi-level feature vector includes user features, operation features, resource features, environmental features, and time-series features; User account attributes and organizational position information are extracted from the user dimension of the access status dataset to form user features. In practice, user account attributes can characterize the access subject's identity category, permission level, and account status, while organizational position information can reflect the access subject's department and position level. The user features formed by combining the two can comprehensively describe the access subject's position in the organizational structure and the context of its access permissions.

[0026] Operational characteristics are formed by extracting operation types, call frequencies, and concurrency from historical operation logs. In practice, operation type indicates the nature of the access request, call frequency reflects the intensity of user requests within a certain timeframe, and concurrency reflects the scale of simultaneous operations. These three types of information can characterize the activity level and potential risk patterns of access behavior, forming operational characteristics for model input.

[0027] The sensitivity level, anonymization requirements, and cross-organizational attributes are extracted from resource sensitivity labels to form resource characteristics. In implementation, the sensitivity level distinguishes the importance of accessing resources, the anonymization requirements specify whether sensitive fields need to be masked or replaced, and the cross-organizational attributes indicate whether the resource is shared across multiple organizations. The combination of these three types of information comprehensively reflects the security attributes of the accessed object, forming resource characteristics.

[0028] By extracting device fingerprints and combining them with network environment credibility, IP reputation score, and geographical location offset, environmental features are formed. In practice, device fingerprints reflect the uniqueness and stability of the accessing terminal, network environment credibility and IP reputation score indicate the risk level of the access source, and geographical location offset indicates the degree of difference between the current access location and historically frequently used locations. The environmental features formed by combining these data can characterize the reliability and anomaly level of the access scenario.

[0029] The system extracts the time period of access, holiday markers, and nighttime access markers to form temporal features. In practice, the time period of access can be used to identify the difference between normal working hours and abnormal hours, while holiday markers and nighttime access markers further supplement the judgment of the reasonableness of access time. The three types of information are combined to form temporal features, which are used to reveal the temporal patterns of access requests.

[0030] The user features, operation features, resource features, environment features, and time features are encoded and concatenated to generate a corresponding state vector. During implementation, each dimension of features is standardized according to a unified field order and encoding rules, and then sequentially concatenated to form a state vector. This state vector fully encompasses information from five aspects: user, operation, resource, environment, and time, and serves as the direct input to the improved neural circuit strategy model.

[0031] In this embodiment, the generation of the access control decision includes: The system receives the processing results of the state vector by the instruction neuron layer of the improved neural circuit strategy model and outputs an initial judgment signal. During implementation, after the state vector enters the instruction neuron layer, the model comprehensively calculates the input features based on its internal connection weights and pre-wiring rules to form a judgment signal representing the risk level of the access request. This judgment signal serves as the basis for subsequent decisions and reflects the initial assessment of the access request's overall security.

[0032] The initial judgment signal is transformed into an access control master decision via the main output unit. This master decision is one of three options: allow, deny, or challenge. In implementation, the main output unit has three mutually exclusive output channels, corresponding to the allow, deny, and challenge decisions, respectively. The model selects one channel to output based on the numerical distribution of the judgment signal, thereby generating a clear access control result. For example, when the judgment signal indicates low risk, it outputs allow; when the risk is too high, it outputs deny; and when the risk is in the intermediate range, it outputs challenge to trigger secondary verification.

[0033] While generating the main decision, fine-grained control parameters are generated by multiple auxiliary output units of the instruction neuron layer. These fine-grained control parameters consist of one or more of the following five types of parameters: Field desensitization parameters are used to mask or obfuscate sensitive fields; Watermark embedding parameters are used to insert traceability identifiers during data export; Rate limiting parameters are used to limit the maximum frequency of access requests; Pagination constraint parameters are used to split large-scale queries into multiple pagination requests; The maximum number of records to export parameter is used to set the maximum number of records that can be exported in a batch. During implementation, the auxiliary output unit runs in parallel with the main output unit, calculating corresponding fine-grained control parameters based on the specific features in the state vector. For example, when an access request involves sensitive fields, the model generates field desensitization parameters to ensure the protection of sensitive information; when an access request involves batch export, the model generates an upper limit parameter for the number of exported records to limit the maximum data volume; when the request frequency is too high, a rate limiting parameter is generated to limit the call rate. In this way, the main decision and fine-grained parameters together form a complete control strategy.

[0034] The master decision and fine-grained control parameters are combined to form an access control decision vector, which is output as the final access control decision result to guide access execution and auditing. In implementation, the master decision and various fine-grained parameters are uniformly encapsulated into a single access control decision vector. This vector contains both the overall admission result and specific execution constraints, enabling it to be directly sent to the gateway layer or database proxy layer for access execution. Simultaneously, it is stored as part of the log, providing a complete chain of evidence for subsequent auditing.

[0035] In this embodiment, the improved neural circuit strategy model includes: The perceptual neuron, comprising multiple input neurons corresponding to user features, operational features, resource features, environmental features, and temporal features, receives component inputs of the state vector. The perceptual neuron layer has multiple input channels, each corresponding to a feature dimension in the state vector; for example, user feature input channels, operational feature input channels, resource feature input channels, environmental feature input channels, and temporal feature input channels. In this way, the model can receive information from different dimensions in parallel and maintain the discriminative power between features during the input phase.

[0036] The perceptual neurons are grouped in a "one-to-five-dimensional correspondence" manner. Each dimension is configured with at least one set of input neurons to carry multiple elements within that dimension (such as permission level, operation type, sensitivity level, device trustworthiness, day / night, etc.). The input neurons within each group perform two functions: first, to receive the necessary elements for that dimension; and second, to receive optional elements within the same dimension. To avoid information interference between dimensions, the input neurons in each group are not directly connected. All cross-dimensional influences are handled by subsequent templated pre-wiring and gating connections, thus achieving the design intent of "distinguishing during the input stage and merging during the connection stage."

[0037] A suppression neuron is configured with several suppression units corresponding to high-risk access conditions. These suppression units are connected to the sensory neuron layer via a fixed pre-wiring method and are used to suppress signals when a preset combination of features is detected. The function of the suppression neuron is to identify combinations of sensitive conditions, such as "late-night access + highly sensitive resources" or "abnormal geographical location + batch export". When the signal output by the sensory neuron meets the above high-risk conditions, the suppression neuron will reduce the transmission strength of the signal or directly block it, thereby preventing high-risk requests from being transmitted to the output end without restriction.

[0038] The inhibitory neuron layer consists of several inhibitory units, each corresponding to one or a set of preset high-risk conditions. Each inhibitory unit structurally includes: Fixed input terminals for specific sensing channels (with fixed connections provided by templated pre-wiring); The aggregation terminal that combines multiple inputs into a single suppressed signal; Connect to the output of the instruction layer and the interpretation layer (the former is used to suppress risks, and the latter is used to record trigger evidence).

[0039] The suppression strength is monotonically related to the input amplitude, and an upper limit is set in the engineering to ensure that hard suppression can be directly formed for "mandatory compliance" scenarios (even if subsequent gating learning cannot exceed the authority), thereby implementing "rigid regulatory constraints" as a strong connection in structure + upper limit of amplitude.

[0040] The instruction neuron, comprising at least three output units, generates allow, deny, and challenge signals, respectively. These are further connected to fine-grained control parameter output units to generate parameters for field anonymization, export restrictions, rate thresholds, and watermarking. The instruction neuron layer is responsible for generating access control results. Its main output units correspond to the three primary decisions: allow, deny, and challenge. Simultaneously, auxiliary output units calculate and output field anonymization parameters, export restriction parameters, rate thresholds, and watermarking parameters based on different input conditions, guiding the platform to implement more refined security measures.

[0041] The instruction neuron layer consists of two parts: the main output sublayer and the auxiliary output sublayer. The main output sublayer carries "allow / reject / challenge" with three mutually exclusive outputs. Its input comes from the combined signals of the perception layer and the inhibition layer. In engineering, a threshold range is set: when the inhibition strength is below the lower threshold, it is biased towards "allow", when it is above the upper threshold, it outputs "reject", and when it is in between, it outputs "challenge".

[0042] The auxiliary output sublayer generates fine-grained control parameters (field desensitization, export restrictions, rate thresholds, and watermark parameters) in parallel channels. These channels directly receive semantically related perceptual inputs and inhibition strengths as adjustment factors, achieving "synchronous shaping of master decision and constraint parameters".

[0043] The relationship between this layer and the innovation lies in the fact that the "strategy" is broken down into discrete main decisions and continuous / discrete constraint parameters, and the structured multi-way output directly supports the effect of "one decision, two outputs".

[0044] The interpreter neuron, comprising several neuronal units connected to inhibitory and sensory neurons, summarizes input contributions and trigger rule identifiers, outputting audit evidence information. The interpreter neuron receives signals from sensory and inhibitory neurons, records which features played a key role in the decision-making process, and outputs the trigger rule identifiers. For example, when a rule leads to access denial, the interpreter neuron outputs the rule number and the main influencing features, thus forming evidence information usable for auditing.

[0045] The interpretation layer consists of two types of units: Rule mapping unit: It corresponds one-to-one with the fixed connection of the templated pre-wired and saves the identifier of the corresponding rule; when the fixed connection is activated and participates in the suppression / decision, the rule mapping unit synchronously outputs the corresponding rule identifier to form "trigger evidence".

[0046] Contribution aggregation unit: Summates and standardizes the activation and inhibition intensities from each sensory channel, outputting the contribution of key features. In engineering, the "product of connection strength and activation amplitude" is used as an intuitive measure of contribution (no formula is needed, only the implementation idea is explained), thereby ensuring that the same access record can provide readable evidence of "which rules were triggered and which features had the greatest impact".

[0047] This layer solidifies interpretability as a structural element: evidence is not post-processed inference, but is generated synchronously by dedicated units within the circuit and output with the log.

[0048] The temporal memory loop establishes a circular connection between the perceptual neuron layer and the inhibitory neuron layer to maintain short-term contextual information of access behavior. The temporal memory loop can record the state information of the user's most recent accesses, so that the model no longer relies on a single input in isolation when judging the current access request, but combines the continuous access history to identify progressive suspicious behavior.

[0049] The temporal memory loop is structurally composed of a back-loop connection and a set of short-time state registers. The back-loop connection sends the intermediate state from the previous (or several) processing iterations back to the current perception-inhibition pathway entry point; the short-time state registers ensure that these states remain updated between multiple requests. In this way, patterns such as "gradually increasing frequency, gradually expanding derived scale, and continuous late-night attempts" are amplified and superimposed in the inhibition layer, making the model inherently sensitive to "progressive risks." This structure makes **temporal context** an intrinsic capability of the circuit, rather than an additional algorithmic module.

[0050] Learnable gating units are inserted into partial connections between perceptual and inhibitory neuron layers, and between inhibitory and instruction neuron layers, to adjust the strength of connection weights during training. These learnable gating units dynamically adjust signal transmission strength; for example, during training, they can enhance signal pathways in normal modes and weaken pathways in abnormal modes through parameter updates. This allows the model to adapt to new access patterns while maintaining prior constraints.

[0051] The gating unit is embedded as an adjustable connection into a "non-mandatory" cross-layer connection (mandatory compliance-type fixed connections are not gating or are only adjustable downwards), and its output is in the form of "input signal × gating coefficient" (the formula is not written, only the principle is explained). Small updates are made during training based on audit feedback. For misplaced samples, increase the gating coefficient of the inhibitory pathway associated with high-risk characteristics or decrease the gating coefficient of the permissive pathway. For samples that are mistakenly blocked, reverse the process to release the normal pathway.

[0052] Meanwhile, to ensure compliance with rigid regulatory constraints, upper and lower limits are set for the gating coefficient, and a strategy of only decreasing and not increasing is adopted for "strongly constrained connections" to ensure that "learning ability" obeys the "prior red line".

[0053] Templated pre-routing consists of a fixed connectivity list, generated based on sources already specified in the claims, such as "resource sensitivity annotations, time characteristics, operation type, geographical location, device trustworthiness, and regulatory rule metadata." In engineering, these prior conditions are mapped to fixed edges of "perception → suppression / perception → instruction." For example, "High-sensitivity resource access at night" is always connected to the strong input of the suppression unit; The "Batch Export ∧ Cross-Organization" function is fixedly connected to the driver of the "Challenge / Limitation Parameters" channel in the instruction layer; The "blacklist rule hit" is always linked to the "reject" channel at the instruction layer.

[0054] These fixed edges are not deleted during the training phase, and their influence strength is constrained to a minimum, ensuring structural compliance and prior interpretability. This is the key innovation of this invention in embedding regulatory rules into the circuit topology.

[0055] In this embodiment, the conflict resolution includes: Access control decisions are compared with pre-defined access control rules in a static policy library. When the results are inconsistent, the mandatory constraints in the static policy library take precedence. The static policy library stores manually pre-defined access control rules, such as prohibiting certain users from accessing classified resources or banning bulk exports during sensitive periods. When access control decisions generated by the improved neural circuit policy model conflict with static rules, the mandatory constraints in the static policy library are executed first to ensure that the hard rules formulated by regulatory authorities are not overridden.

[0056] The access control decisions are compared with the regulatory red-line rules in the compliance module. When a violation of mandatory compliance requirements is found, the access control decision is adjusted to deny or challenge. The compliance module contains industry-level compliance requirements, such as the prohibition of cross-border transmission of sensitive data and the prohibition of displaying classified information on unauthorized terminals. When an access control decision triggers these red-line rules, the original result is automatically overridden and adjusted to deny access, or additional verification is required in some cases, thereby ensuring that access behavior complies with mandatory regulations.

[0057] Access control decisions are validated using a spatiotemporal quota ledger. This ledger records available access quotas indexed by resource identifiers and time slices. When an access request exceeds the remaining quota for the corresponding resource, the access control decision is downgraded to pagination, rate limiting, or direct rejection. The spatiotemporal quota ledger manages the usage quota of accessed resources within a specific time period; for example, it specifies a maximum number of records that can be exported from a highly sensitive resource within one minute. When the access decision generated by the model exceeds the quota limit, the decision is downgraded, such as by forcing pagination, limiting access frequency, or direct rejection, thereby preventing large-scale sensitive data leakage in a short period.

[0058] After completing the comparison, verification, and limit validation, a composite access control result with conflict resolution is output. The results from these three aspects are combined to form the final composite access control result. This result reflects the intelligent model's judgment, meets static rules and compliance requirements, and is protected by limit constraints, ensuring that access control is both flexible and compliant with mandatory regulatory requirements.

[0059] In this embodiment, the generation of the audit log includes: When executing access control decisions after conflict resolution, the user identifier, resource identifier, operation type, and execution time of the access request are recorded. Simultaneously, a log record is generated when access control is actually executed at the gateway or database proxy layer. This record contains the user identifier to identify the accessing subject, the resource identifier to identify the accessed object, the operation type to describe the access behavior, and the execution time to determine the specific point in time when the access occurred. This information constitutes the basic content of the audit log.

[0060] The improved neural circuit strategy model's interpretive neuron layer is invoked to extract the triggered access control rules, key feature contributions, and quota ledger status, outputting them as evidence information. The interpretive neuron layer outputs the decision basis for the access request, including the triggered access control rule number, the input features that play a major role in the final result and their contributions, and the remaining state of the access quota. In this way, evidence information associated with each decision can be generated, ensuring that the results are not only usable but also traceable.

[0061] The access request information is combined with the evidence information to generate a complete audit log entry. The basic information of the access request and the evidence information are integrated into a single complete log entry, thus simultaneously preserving the decision-making process and the execution result. This entry reflects both the access behavior and the rationale behind the decision, ensuring the completeness of the audit materials.

[0062] The audit log entries are stored in an audit log repository. Complete audit log entries are stored in a dedicated audit log repository and archived in chronological order. This ensures long-term data traceability and provides a reliable data source for subsequent manual verification, compliance checks, and model retraining.

[0063] In this embodiment, the online fine-tuning and offline retraining include: Based on the audit logs and manual review results, execution feedback of access control decisions is extracted, and cases of erroneous granting and blocking are marked. Decision execution status is periodically extracted from the audit logs, such as whether access requests were allowed, denied, or challenged, and erroneous granting (risky requests incorrectly allowed) and erroneous blocking (legitimate requests incorrectly denied) are marked in conjunction with manual review results. These feedback labels provide accurate training signals for subsequent model optimization.

[0064] During the online operation phase, the learnable gating units of the improved neural circuit strategy model are fine-tuned using a small-step parameter update method to adapt to the latest access behavior patterns. The model is kept running continuously, and the connection weights of the learnable gating units are slightly adjusted without affecting online services. The purpose of this fine-tuning is to enable the model to adapt promptly to recent changes in access behavior trends, such as new access frequency patterns or new abnormal combination features, thereby improving the accuracy of real-time judgment.

[0065] During the offline training phase, newly added audit log samples and feedback labels are incorporated into the training set, and the improved neural circuit strategy model is retrained using historical data to update connection weights and rule mappings. Audit logs accumulated over a period of time and their corresponding feedback labels are merged with the original training dataset to form an expanded training set. Retraining the model offline comprehensively optimizes the connection weights and rule mappings between neurons, ensuring the model maintains long-term stable recognition capabilities.

[0066] After updating the improved neural circuit strategy model, a verification process is performed to ensure that the output of the improved neural circuit strategy model does not violate the mandatory constraints of the compliance module. The online model is then gradually replaced through a canary release approach. After confirming that there are no errors, the online model is gradually replaced using a canary release approach, that is, the new model is first enabled in some request scenarios, and the application scope is gradually expanded to achieve a smooth transition. Example

[0067] To verify the feasibility of this invention in practice, it was applied to a production-equivalent test environment of a provincial tobacco regulatory platform (using the same data source, network, and hardware specifications as the official production environment). The test period was 30 days, based on the platform's historical audit logs and current actual access requests, and was implemented according to the following steps: Data collected included 1,268 platform users, 312 types of controlled resources (74 of which were highly sensitive), and 1,203,417 access requests.

[0068] Data processing: Following the process of "unified timestamp alignment - quality verification - five-dimensional organization - concatenated index" of this invention, an access status dataset is formed; then, multi-level features are extracted based on the five dimensions of "user / operation / resource / environment / time" and encoded into a status vector.

[0069] Training / validation segmentation: The "mistakenly released / mistakenly blocked / violation / compliance" samples marked in the historical audit logs are used as monitoring signals; the first 21 days are used for training and parameter selection, and the last 9 days are used for online gray-scale validation.

[0070] Structure: Composed of sensory neurons, inhibitory neurons, instruction neurons, and interpretation neurons; it features a templated pre-wired structure, learnable gating units, and a time memory loop.

[0071] Gray scale: The gateway layer is divided into 50% / 50% user hashes: The control group uses "static policy library + compliance module + quota ledger" (RBAC / ABAC baseline, denoted as Baseline); the experimental group uses "access control decision output by the model of this invention + conflict resolution", denoted as Proposed.

[0072] Compliance verification: Before the gray-scale rollout, four layers of verification are completed: "rule consistency, boundary scenarios, historical log playback, compliance signature and archiving", and continuous monitoring and rollback are possible during the gray-scale rollout.

[0073] The evaluation indicators include: Misplacement rate (%): The percentage of cases that were initially allowed but were subsequently deemed non-compliant / high-risk by audits.

[0074] False blocking rate (%): The percentage of cases that were initially rejected / challenged but were subsequently deemed valid by an audit.

[0075] Master decision accuracy rate (%): The percentage of master decisions that are consistent with the final audit decision.

[0076] Audit evidence completeness rate (%): Whether each decision simultaneously generates "trigger rule identifier + key feature contribution + quota ledger status".

[0077] Average decision delay (ms): The average time taken from the state vector to the access control decision vector output (excluding network round trips).

[0078] Compliance module conflict coverage (%): The percentage of model outputs that are successfully corrected to compliant results when they conflict with mandatory compliance.

[0079] Pagination constraint / rate limit trigger count, export limit interception count, quota overrun and downgrade count: actual trigger counts of fine-grained control parameters.

[0080] Table 1 Comparison Results Chart ; As shown in Table 1, the method of this invention outperforms the baseline method in terms of false release rate, false blocking rate, and master decision accuracy. The false release rate is reduced by 56.1%, the false blocking rate is reduced by 43.6%, and the master decision accuracy is improved by 2.69 percentage points. At the same time, the audit evidence completeness rate is significantly improved to 96.3%, and the compliance conflict coverage rate reaches 100%. This indicates that the present invention, through structural innovations such as templated pre-wiring, time memory loop, and learnable gating units, achieves effective suppression of risky access and comprehensive coverage of compliance constraints, thereby enhancing the accuracy and interpretability of access control.

[0081] In terms of fine-grained control, the instruction neuron layer of this invention can synchronously generate parameters such as pagination constraints, rate limits, and export limits, significantly increasing the number of triggers for pagination, rate limiting, and degradation execution. This strengthens risk management without increasing excessive rejections, ensuring a balance between business continuity and security. Although the average decision latency only increases by about 2.3 milliseconds, it is completely acceptable within the gateway layer. Therefore, this invention significantly improves the refinement and practicality of access control for tobacco regulatory platforms while ensuring real-time performance.

[0082] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based access control method for a tobacco regulatory platform, characterized in that, Includes the following steps: Collect user access data, align it with a unified timestamp, and form an access status dataset; A multi-level feature vector is constructed based on the access state dataset, and the multi-level feature vector is encoded into a state vector; The state vector is input into the improved neural circuit strategy model, which consists of sensory neurons, inhibitory neurons, instruction neurons and interpreting neurons. It has a templated pre-wired structure, learnable gating units and time memory loops, and is used to output access control decisions. The access control decision is reconciled with the static policy library, the mandatory compliance module and the time-space quota ledger. If there is a conflict, compliance and quota constraints are executed first to form a composite access control result. Based on the synthesized access control results, access operations are performed at the gateway layer or database proxy layer of the tobacco regulatory platform, and audit logs carrying evidence information are generated. Based on the audit logs and manual verification results, the improved neural circuit strategy model is fine-tuned online or retrained offline.

2. The access control method for a tobacco regulatory platform based on machine learning according to claim 1, characterized in that, The formation of the access state dataset includes: The accessed user data includes user account attributes, organizational position information, historical operation logs, device fingerprints, network environment, geographical location, time characteristics, resource sensitivity annotations, and regulatory rule metadata. Align data entries from different sources according to a globally unified timestamp; Perform quality checks on the collected data, removing duplicate logs, invalid fields, and obvious outliers; The cleaned data entries are organized into five dimensions: user, operation, resource, environment, and time; Using "user ID - access timestamp - operation ID" as a combined index, the data from the above five dimensions are concatenated into an access status record. All records are stored in chronological order to form an access status dataset.

3. The access control method for a tobacco regulatory platform based on machine learning according to claim 1, characterized in that, The generation of the state vector includes: The multi-level feature vector includes user features, operation features, resource features, environmental features, and time-series features; User account attributes and organizational job information are extracted from the user dimension of the access status dataset to form user features; Operation characteristics are formed by extracting operation types, call frequencies, and concurrency from historical operation logs from an operational perspective. The sensitivity level, desensitization requirements, and cross-institutional attributes in resource sensitivity labeling are extracted from the resource dimension to form resource features; From an environmental perspective, device fingerprints are extracted along with trustworthiness, IP reputation score, and geographical location offset within the network environment to form environmental characteristics; Extract the time period, holiday markers, and nighttime access markers from the time dimension to form time features; The user features, operation features, resource features, environmental features, and time features are encoded and concatenated to generate the corresponding state vector.

4. The access control method for a tobacco regulatory platform based on machine learning according to claim 1, characterized in that, The generation of the access control decision includes: The system receives the processing results of the state vector by the instruction neuron layer of the improved neural circuit strategy model and outputs the initial judgment signal. The initial determination signal is converted into an access control master decision via the main output unit, and the master decision is one of three options: allow, deny, or challenge. While generating the main decision, fine-grained control parameters are generated by multiple auxiliary output units of the instruction neuron layer. These fine-grained control parameters consist of one or more of the following five types of parameters: Field desensitization parameters are used to mask or obfuscate sensitive fields; Watermark embedding parameters are used to insert traceability identifiers during data export; Rate limiting parameters are used to limit the maximum frequency of access requests; Pagination constraint parameters are used to split large-scale queries into multiple pagination requests; The maximum number of records to export parameter is used to set the maximum number of records that can be exported in a batch. The master decision is combined with the fine-grained control parameters to form an access control decision vector, which is then output as the final access control decision result to guide access execution and auditing.

5. The access control method for a tobacco regulatory platform based on machine learning according to claim 1, characterized in that, The improved neural circuit strategy model includes: The sensory neuron includes multiple input neurons, which correspond to user features, operation features, resource features, environmental features and time features, respectively, and are used to receive component inputs of the state vector; Suppressing neurons are configured with several suppression units corresponding to high-risk access conditions. The suppression units are connected to the perceptual neuron layer through a fixed pre-wiring method and are used to suppress the signal when a preset feature combination is detected. The instruction neuron includes at least three output units, which are used to generate allow signals, rejection signals and challenge signals, respectively, and are further connected to fine-grained control parameter output units to generate field desensitization, export restrictions, rate thresholds and watermark parameters; The explanatory neuron, which includes several neuronal units connected to inhibitory neurons and sensory neurons, is used to summarize the input contribution and trigger rule identifiers and output audit evidence information; Temporal memory loops establish circular connections between the perceptual neuron layer and the inhibitory neuron layer to maintain short-term contextual information of access behavior; Learnable gating units are inserted into partial connections between the perceptual neuron layer and the inhibitory neuron layer, and between the inhibitory neuron layer and the instruction neuron layer, to adjust the connection weight strength during the training phase.

6. The access control method for a tobacco regulatory platform based on machine learning according to claim 1, characterized in that, The conflict resolution includes: The access control decision is compared with the preset access control rules in the static policy library. When the results are inconsistent, the mandatory constraint rules in the static policy library shall be given priority. The access control decision after comparison is checked against the regulatory red line rules in the compliance module. When a violation of mandatory compliance requirements occurs, the access control decision is adjusted to deny or challenge. Access control decisions are validated based on a spatiotemporal quota ledger. The spatiotemporal quota ledger records available access quotas using resource identifiers and time slices as indexes. When an access request is detected to exceed the remaining quota of the corresponding resource, the access control decision is downgraded to pagination execution, rate limiting execution, or direct rejection. After completing the comparison, verification, and quota check, the synthesized access control result with conflict resolution is output.

7. The access control method for a tobacco regulatory platform based on machine learning according to claim 1, characterized in that, The generation of the audit log includes: When implementing access control decisions after conflict resolution, the user identifier, resource identifier, operation type, and execution time of the access request are recorded; The interpretation neuron layer of the improved neural circuit strategy model is invoked to extract the triggered access control rules, key feature contribution, and quota ledger status, which are then output as evidence information. The access request information is combined with the evidence information to generate a complete audit log entry; The audit log entries are stored in the audit log database.

8. The access control method for a tobacco regulatory platform based on machine learning according to claim 1, characterized in that, The online fine-tuning and offline retraining include: Based on the audit logs and manual verification results, the execution feedback of access control decisions is extracted, and cases of accidental access and accidental blocking are marked. During the online operation phase, the learnable gating units of the improved neural circuit strategy model are fine-tuned using a small-step parameter update method to adapt to the latest access behavior patterns. During the offline training phase, newly added audit log samples and feedback labels are included in the training set, and the improved neural circuit strategy model is retrained in combination with historical data to update connection weights and rule mappings. After the improved neural circuit strategy model is updated, a verification is performed to ensure that the output of the improved neural circuit strategy model does not violate the mandatory constraints of the compliance module, and the online model is gradually replaced through a canary release method.

Images