Electronic government affair service method and device based on generative artificial intelligence

Abstract

The invention discloses an electronic government affair service method and device based on generative artificial intelligence, belongs to the technical field of government affair services, and aims to solve the technical problems of how to realize government affair service intelligence and provide more convenient and efficient government affair service experience for the public. According to the technical scheme, the method comprises the following steps: constructing a standardized data carrying out multi-dimensional feature extraction on a license and a government affair text through a sensitive semantic recognition engine, and realizing data full-process traceability in combination with a data ownership block chain evidence storage system to form a standardized and traceable government affair data asset pool; establishing a credible decision center: constructing an interpretable neural network architecture according to a standardized data base, dynamically analyzing a basis and a data dependency item in an examination and approval decision through a causal inference map, introducing an adversarial generative network sandbox test to perform a pressure test on an algorithm, detecting model robustness in a data poisoning and injection risk prompting scene, and establishing a credible decision center; obtaining a decision track and a running log; and constructing a safety guarantee system.

Description

Technical Field

[0001] The present invention relates to the field of government service technology, and specifically to a generative artificial intelligence-based e-government service method and device. Background Art

[0002] While the application of artificial intelligence in government services has brought significant changes, it also carries a range of risks. On the one hand, the rapid development of technology may lag behind regulatory frameworks. On the other hand, the automated and intelligent nature of algorithms may shift power in decision-making to algorithms or systems. Furthermore, the risk of government or public data leakage cannot be ignored, making the security and privacy of this data a major challenge.

[0003] The core issue of the systemic contradictions caused by the integration of generative artificial intelligence (AIGC) technology into the e-government service system can be summarized as the structural conflict between technological empowerment and public governance, which manifests itself in three dimensions of governance dilemma:

[0004] First, there's an imbalance between technological autonomy and administrative controllability. While generative AI improves government service efficiency, its black-box nature renders its decision-making logic untraceable. Furthermore, the expansion of technological power and the lag in regulatory frameworks create risks that are difficult to effectively contain.

[0005] Second, data governance conflicts with public interests. The multi-source, heterogeneous nature of government data (such as license information and policy texts) leads to uneven training data quality. Data barriers between departments result in insufficient supply of effective corpus, while excessive data mining may infringe on citizens' privacy.

[0006] Third, there is a contradiction between the accessibility of technology and the fairness of services. When optimizing the service experience, generative AI may exacerbate the digital divide - the intelligent question-answering system lacks dialect recognition, and the visual interface relies on digital devices, resulting in a decrease in the accessibility of services for vulnerable groups. A more hidden risk is that algorithmic bias may lead to implicit discrimination. For example, the policy recommendation system ignores the demands of special groups due to training data bias. The essence of these problems is the fundamental contradiction between the high efficiency demands of intelligent technology and the fairness and security requirements of public services. The present invention is committed to establishing a balance between technology empowerment and risk prevention and control by constructing a classified supervision system (data hierarchical review and algorithm transparency), a collaborative governance mechanism (government-enterprise collaboration and public participation), and a dynamic risk assessment framework, and ultimately realizing the transformation of government services from "efficiency first" to the "efficiency-safety-fairness" trinity governance paradigm.

[0007] Therefore, how to realize intelligent government services and provide the public with a more convenient and efficient government service experience is a technical problem that needs to be solved urgently. Summary of the Invention

[0008] The technical task of the present invention is to provide a method and device for electronic government services based on generative artificial intelligence to solve the problem of how to realize intelligent government services and provide the public with a more convenient and efficient government service experience.

[0009] The technical task of the present invention is achieved in the following manner: a generative artificial intelligence e-government service method, the method is as follows:

[0010] Build a standardized data base: Use a sensitive semantic recognition engine to extract multi-dimensional features from certificates and government documents, and combine it with a blockchain-based data ownership system to achieve full data traceability, forming a standardized and traceable government data asset pool.

[0011] Establishing a trusted decision-making hub: Build an interpretable neural network architecture based on a standardized data foundation. Dynamically analyze the basis and data dependencies in approval decisions through causal inference graphs. Stress-test the algorithm using a generative adversarial network sandbox test to detect model robustness in data poisoning and prompt injection risk scenarios, thereby obtaining decision traces and operation logs. Data poisoning is an attack method against machine learning models that, by inserting malicious samples into training data or tampering with data labels, induces the model to learn incorrect patterns, causing it to output the erroneous results expected by the attacker after deployment.

[0012] Build a security assurance system: Based on decision-making trajectories and operation logs, a spatiotemporal feature analysis model is constructed to monitor abnormal data flows in real time. Differential privacy dynamic desensitization technology is used to implement quantum encryption for important data, and homomorphic encryption is enabled for special data to form a hierarchical protection system. At the same time, a service modality adaptation engine is deployed to automatically switch interaction modes based on user portraits, and cross-departmental risk joint prevention and control is achieved through a federal audit mechanism.

[0013] Preferably, the standardized data base has the following:

[0014] The sensitive semantic recognition engine uses a pre-trained language model (such as RoBERTa) and a graph convolutional network fusion architecture to perform ternary classification of government data on certificates and government documents, enabling intelligent classification of government data. It also uses transfer learning to adapt to vertical domain terminology libraries to build over 300 sensitive entity recognition rules, ensuring an accuracy rate of over 98% for identifying important data.

[0015] Dynamic mapping of cross-departmental data contribution and equity distribution: Implement intelligent measurement and equity distribution of each department's data contribution through the data ownership blockchain notarization system;

[0016] Multi-source heterogeneous data fusion based on federated learning: Deploy a horizontal federated learning system, where each government node retains local data, achieves cross-departmental feature fusion through encrypted parameter exchange, and uses differential privacy technology to ensure that data is "available but invisible", solving the dilemma of data silos while preventing privacy leaks.

[0017] Preferably, the three-category classification of government data includes important data, special data, and general data; among them, important data is subject to full-process manual review and quantum encryption storage; special data adopts a dynamic authorization access mechanism; and general data is open for sharing after desensitizing processing.

[0018] More optimally, the data ownership blockchain notarization system specifically includes: building a government alliance chain based on Hyperledger Fabric, realizing the chain notarization of the entire process of data collection, labeling, and sharing, and developing a smart contract to automatically execute a data contribution measurement model, establishing a "data contribution-equity distribution" dynamic mapping mechanism, and eliminating barriers to data sharing between departments.

[0019] As a preferred option, the trusted decision-making center is established as follows:

[0020] Interpretable neural network architecture: Builds a GNN-based causal inference model, automatically generates a three-dimensional map of the decision path, and uses attention weight visualization technology to intuitively display the basis, data dependencies, and weight ratios of approval decisions, thereby increasing algorithm tracing efficiency by three times.

[0021] Adversarial Generative Network Sandbox Testing: We deployed an algorithm stress testing environment, used GAN to generate extreme case simulation systems, and conducted adversarial attack tests on government approval scenarios. We tested the model's robustness against data poisoning and prompt injection attacks, discovering and fixing 23% of potential decision-making vulnerabilities.

[0022] Reinforcement learning parameter optimization: Design a multi-objective reward function that integrates service efficiency and fairness indicators (such as response time for digitally disadvantaged groups), and dynamically adjust algorithm parameters through the DQN algorithm to achieve an automatic balance between technology empowerment and risk prevention and control.

[0023] As a preferred option, the security assurance system is constructed as follows:

[0024] Time-series graph neural network analysis: This integrates data access logs and decision trajectory data to build a spatiotemporal feature analysis model, and detects abnormal flow patterns of sensitive data (such as high-frequency cross-domain access) in real time. Risk assessment and triggering of circuit breakers are completed within 200ms, significantly reducing the risk of data leakage.

[0025] Differential privacy dynamic desensitization: Attribute-based encryption (ABE) gateways are deployed based on data classification labels. Important data is encrypted using quantum key distribution, special data is processed using homomorphic encryption, and general data is protected by a k-anonymization threshold, achieving a dynamic balance between hierarchical protection and business efficiency.

[0026] Service modality adaptive engine: Builds a user-profile-driven intelligent dispatching system; automatically switches the voice interaction channel and increases font contrast when an elderly user is detected; and intelligently dispatches offline service outlets to initiate emergency responses when the regional network failure rate exceeds a threshold, ensuring 100% service coverage.

[0027] A generative artificial intelligence-based e-government service device, comprising:

[0028] The data governance layer is used to extract multi-dimensional features from certificates and government documents through a sensitive semantic recognition engine, and to achieve full data traceability through a blockchain-based data ownership system, thereby forming a standardized and traceable government data asset pool.

[0029] The algorithm transparency collaboration layer is used to build an interpretable neural network architecture based on a standardized data base. It dynamically analyzes the basis and data dependencies in approval decisions through causal inference graphs. It also introduces adversarial generative network sandbox testing to stress-test the algorithm, detect model robustness under data poisoning and prompt injection risk scenarios, and obtain decision trajectories and operation logs. Data poisoning is an attack method against machine learning models that inserts malicious samples into training data or tampers with data labels to induce the model to learn incorrect patterns, causing it to output the incorrect results expected by the attacker after deployment.

[0030] The prevention and control layer is used to build a spatiotemporal feature analysis model based on decision trajectories and operation logs to monitor abnormal data flows in real time, and adopt differential privacy dynamic desensitization technology to implement quantum encryption for important data, and enable homomorphic encryption for special data to form a hierarchical protection system; at the same time, a service modality adaptation engine is deployed to automatically switch the interaction mode according to the user portrait, and realize cross-departmental risk joint prevention and control through the federal audit mechanism.

[0031] Preferably, the data governance layer includes:

[0032] The sensitive semantic recognition engine is used to perform ternary classification of government data on certificates and government documents using a pre-trained language model (such as RoBERTa) and a graph convolutional network fusion architecture, achieving intelligent classification of government data. It also uses transfer learning to adapt to vertical domain terminology libraries to build 300+ sensitive entity recognition rules, ensuring an accuracy rate of over 98% for identifying important data. The ternary classification of government data includes important data, special data, and general data. Important data undergoes full-process manual review and quantum encryption storage; special data uses a dynamic authorization access mechanism; and general data is open for sharing after being desensitized.

[0033] The data ownership blockchain evidence storage system is used to build a government alliance chain based on Hyperledger Fabric, realize the full process of data collection, labeling, and sharing, and store it on the chain. It also develops a smart contract to automatically execute the data contribution measurement model, establishes a dynamic mapping mechanism of "data contribution-equity distribution", and eliminates barriers to data sharing between departments.

[0034] The federated learning system is used to deploy a horizontal federated learning system. Each government node retains local data, achieves cross-departmental feature fusion through encrypted parameter exchange, and adopts differential privacy technology to ensure that data is "available but invisible", breaking the dilemma of data silos while preventing privacy leakage risks.

[0035] Preferably, the algorithm transparency collaboration layer includes:

[0036] An interpretable neural network architecture unit is used to build a GNN-based causal inference model, automatically generate a three-dimensional map of the decision path, and intuitively display the basis, data dependencies, and weight ratios of the approval decision through Attention weight visualization technology, thereby increasing the algorithm traceability efficiency by three times.

[0037] The Generative Adversarial Network Sandbox Testing Unit deploys an algorithm stress testing environment, uses GAN to generate extreme case simulation systems, and conducts adversarial attack tests on government approval scenarios. This tests the model's robustness against data poisoning and prompt injection attacks, discovering and fixing 23% of potential decision-making vulnerabilities.

[0038] The reinforcement learning parameter optimization unit is used to design a multi-objective reward function, integrate service efficiency and fairness indicators (such as the response time of digitally disadvantaged groups), and dynamically adjust algorithm parameters through the DQN algorithm to achieve an automatic balance between technology empowerment and risk prevention and control.

[0039] Preferably, the prevention and control layer includes:

[0040] The time-series graph neural network analysis unit integrates data access logs and decision trajectory data to build a spatiotemporal feature analysis model. It also detects abnormal flow patterns of sensitive data (such as high-frequency cross-domain access) in real time, completes risk classification and triggers the circuit breaker mechanism within 200ms, significantly reducing the risk of data leakage.

[0041] A differential privacy dynamic desensitization unit is used to deploy an attribute-based encryption (ABE) gateway based on data classification labels. Important data is encrypted using quantum key distribution, special data is processed using homomorphic encryption, and general data is set with a k-anonymization threshold, achieving a dynamic balance between hierarchical protection and business efficiency.

[0042] The service modality adaptation engine is used to build a user-profile-driven intelligent dispatching system. When an elderly user is detected, the system automatically switches the voice interaction channel and increases the font contrast. When the regional network failure rate exceeds the threshold, the intelligent dispatch offline service outlets initiate an emergency response to ensure that the service coverage rate remains 100%.

[0043] The generative artificial intelligence-based e-government service method and device of the present invention have the following advantages:

[0044] (1) This invention systematically resolves the structural contradictions between AI technology and government services by constructing a three-stage progressive system of "data hierarchical governance - algorithm transparency and collaboration - full-process prevention and control". The data governance layer establishes a standardized data foundation, the algorithm layer builds a trusted decision-making center, and the prevention and control layer forms a closed loop of security assurance. The technical achievements of the three stages are transmitted step by step along the path of "foundation construction → intelligent application → risk mitigation", ultimately achieving a dynamic balance in the efficiency, security, and fairness of the government system;

[0045] (2) This invention systematically resolves the core contradictions in the integration of generative AI and e-government by constructing a three-stage progressive technical architecture. Regarding the imbalance between technological autonomy and administrative controllability, the decision path visualization module constructed based on an interpretable neural network fully presents the algorithm decision logic through a causal inference graph, making the decision basis of automated approval scenarios traceable and reviewable. At the same time, relying on the algorithm sandbox jointly built by the government and enterprises, it implements stress testing on key decision nodes to ensure that the human-machine joint signature mechanism is automatically triggered when the degree of AI intervention exceeds the threshold. This not only retains the advantage of technological autonomy, but also maintains the statutory discretion of administrative entities through a responsibility tracing system, effectively eliminating the trust crisis caused by the algorithm black box.

[0046] (3) This invention addresses the conflict between data governance and public interests. The three-element hierarchical architecture of government data uses a sensitive semantic recognition engine to achieve intelligent classification of data such as certificates and policy texts. It combines the blockchain evidence storage system to build a data ownership traceability mechanism, which significantly improves the manual review rate of important data throughout the entire process and effectively reduces the risk of illegal access to special data. The standardized data asset pool output by the data governance layer uses dynamic desensitization technology to achieve compliant integration of social security data and credit data, ensuring the quality of training data while controlling the risk of privacy leakage, successfully solving the value paradox of data openness and sharing and privacy protection.

[0047] (IV) In terms of technology accessibility and service fairness, the risk warning matrix constructed by the prevention and control layer dynamically adjusts the service supply model by real-time monitoring of fairness indicators such as dialect interaction failure rate and digital device dependence; the digital inclusion module combines voice recognition enhancement technology to improve the interactive experience of the elderly group, and the offline service redundancy mechanism established simultaneously ensures government coverage in remote areas; the algorithm bias correction model is used to conduct feature reinforcement learning on the service demands of special groups to eliminate implicit discrimination caused by data bias. The resulting intelligent government ecosystem achieves continuous optimization of government service satisfaction and significantly improves the accessibility of services for vulnerable groups. The technical achievements at all levels form a closed-loop transmission mechanism: data classification standards lay the foundation for algorithm transparency, and the decision logic output by the explainable architecture provides a basis for parameter tuning for risk prevention and control. The dynamically optimized service model feeds back into the iteration of data governance strategies. This cyclical evolution of "standard setting-collaborative governance-risk prevention and control" promotes the e-government system from single-dimensional efficiency priority to a "efficiency-security-fairness" trinity governance paradigm;

[0048] (5) The classification accuracy of the data governance layer of the present invention determines the depth of algorithm interpretability, the decision transparency of the algorithm layer affects the accuracy of risk identification, and the feedback data of the risk prevention and control layer reversely optimizes the data desensitization strategy. The three stages realize two-way data flow through the standard API interface, forming an enhanced loop of "the more accurate the data - the more reliable the decision - the smarter the protection", and building a safe, controllable and continuously evolving intelligent government ecosystem. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below with reference to the accompanying drawings.

[0050] Attachment Figure 1 This is a flowchart of the e-government service method based on generative artificial intelligence. DETAILED DESCRIPTION

[0051] The following detailed description of the generative artificial intelligence-based e-government service method and device of the present invention is given with reference to the accompanying drawings and specific embodiments.

[0052] Example 1:

[0053] As attached Figure 1 As shown, this embodiment provides a generative artificial intelligence e-government service method, which is specifically as follows:

[0054] S1. Build a standardized data base: Use a sensitive semantic recognition engine to extract multi-dimensional features from certificates and government documents, and combine this with a blockchain-based data ownership system to achieve full data traceability, forming a standardized and traceable government data asset pool.

[0055] S2. Establish a trusted decision-making center: Construct an interpretable neural network architecture based on a standardized data base, dynamically analyze the basis and data dependencies in approval decisions through causal inference graphs, and introduce adversarial generative network sandbox testing to stress test the algorithm, detect data poisoning and the robustness of the model under risk injection scenarios, and then obtain decision trajectories and operation logs. Among them, data poisoning is an attack method against machine learning models. By implanting malicious samples in training data or tampering with data labels, the model is induced to learn incorrect patterns, causing it to output the erroneous results expected by the attacker after deployment. The core achievements of this stage are: ① A three-dimensional visualization map of the decision path to achieve full traceability of the algorithm logic; ② A human-machine joint signing mechanism dynamically balances AI autonomy and administrative discretion; ③ A reinforcement learning-driven parameter optimization model embeds service efficiency and fairness indicators into the algorithm iteration process. These achievements enable automated decision-making to retain technical advantages while meeting the compliance requirements of public governance.

[0056] S3. Build a security assurance system: Based on decision trajectories and operation logs, we build a spatiotemporal feature analysis model to monitor abnormal data flows in real time. We also employ differential privacy dynamic desensitization technology to implement quantum encryption for important data and enable homomorphic encryption for special data, forming a hierarchical protection system. We also deploy a service modality adaptation engine to automatically switch interaction modes based on user profiles, and implement cross-departmental risk prevention and control through a federal audit mechanism. Ultimately, we develop two core capabilities: 1. Response time for data breach risks is reduced to seconds; 2. A dynamic balance is achieved between bridging the digital divide and ensuring service fairness.

[0057] The standardized data base in step S1 of this embodiment has the following:

[0058] S101, the sensitive semantic recognition engine uses a pre-trained language model (such as RoBERTa) and a graph convolutional network fusion architecture to perform ternary classification of government data on certificates and government documents, achieving intelligent classification of government data. It also uses transfer learning to adapt to vertical domain terminology libraries to build more than 300 sensitive entity recognition rules, ensuring an accuracy rate of over 98% for important data recognition.

[0059] S102. Dynamic mapping of cross-departmental data contribution and equity distribution: Implement intelligent measurement and equity distribution of each department’s data contribution through the data ownership blockchain notarization system;

[0060] S103. Multi-source heterogeneous data fusion based on federated learning: Deploy a horizontal federated learning system. Each government node retains local data, achieves cross-departmental feature fusion through encrypted parameter exchange, and adopts differential privacy technology to ensure that data is "available but invisible", solving the dilemma of data silos while preventing privacy leakage risks.

[0061] The ternary classification of government data in this embodiment includes important data, special data and general data; among them, important data is subject to full-process manual review and quantum encryption storage; special data adopts a dynamic authorization access mechanism; general data is open for sharing after desensitizing processing.

[0062] The data ownership blockchain notarization system in this embodiment specifically builds a government alliance chain based on Hyperledger Fabric to realize the on-chain notarization of the entire process of data collection, labeling, and sharing, and develops a smart contract to automatically execute a data contribution measurement model, establishes a dynamic mapping mechanism of "data contribution-equity distribution", and eliminates barriers to data sharing between departments.

[0063] The specific steps for establishing a trusted decision-making center in step S2 of this embodiment are as follows:

[0064] S201, Interpretable Neural Network Architecture: Builds a GNN-based causal inference model, automatically generates a three-dimensional map of the decision path, and uses Attention weight visualization technology to intuitively display the basis, data dependencies, and weight ratios in the approval decision, thereby increasing algorithm traceability efficiency by three times.

[0065] S202, Generative Adversarial Network Sandbox Testing: Deploy an algorithm stress testing environment, use GAN to generate extreme case simulation systems, and conduct adversarial attack tests on government approval scenarios. This test examines the model's robustness against data poisoning and prompt injection attacks, discovering and fixing 23% of potential decision-making vulnerabilities.

[0066] S203, reinforcement learning parameter optimization: Design a multi-objective reward function, integrate service efficiency and fairness indicators (such as response time for digitally disadvantaged groups), and dynamically adjust algorithm parameters through the DQN algorithm to achieve an automatic balance between technology empowerment and risk prevention and control.

[0067] The construction of the security assurance system in step S3 of this embodiment is specifically as follows:

[0068] S301, Time Series Graph Neural Network Analysis: This integrates data access logs and decision trajectory data to build a spatiotemporal feature analysis model, and detects abnormal flow patterns of sensitive data (such as high-frequency cross-domain access) in real time. Risk assessment and triggering of circuit breakers are completed within 200ms, significantly reducing the risk of data leakage.

[0069] S302, Differential Privacy Dynamic Desensitization: Deploy an Attribute-Based Encryption (ABE) gateway based on data classification labels. Important data is encrypted using quantum key distribution, special data is processed using homomorphic encryption, and general data is set with a k-anonymization threshold, achieving a dynamic balance between hierarchical protection and business efficiency.

[0070] S303, Service Mode Adaptation Engine: Builds a user-profile-driven intelligent dispatching system; automatically switches the voice interaction channel and increases the font contrast when an elderly user is detected; when the regional network failure rate exceeds the threshold, the intelligent dispatch offline service outlets initiate an emergency response to ensure that the service coverage rate remains at 100%.

[0071] Example 2:

[0072] This embodiment provides a generative artificial intelligence-based e-government service device, which includes:

[0073] The data governance layer is used to extract multi-dimensional features from certificates and government documents through a sensitive semantic recognition engine, and to achieve full data traceability through a blockchain-based data ownership system, thereby forming a standardized and traceable government data asset pool.

[0074] The algorithm transparency collaboration layer is used to build an interpretable neural network architecture based on a standardized data base. It dynamically analyzes the basis and data dependencies in approval decisions through causal inference graphs. It also introduces adversarial generative network sandbox testing to stress-test the algorithm, detect model robustness under data poisoning and prompt injection risk scenarios, and obtain decision trajectories and operation logs. Data poisoning is an attack method against machine learning models that inserts malicious samples into training data or tampers with data labels to induce the model to learn incorrect patterns, causing it to output the incorrect results expected by the attacker after deployment.

[0075] The prevention and control layer is used to build a spatiotemporal feature analysis model based on decision trajectories and operation logs to monitor abnormal data flows in real time. It also uses differential privacy dynamic desensitization technology to implement quantum encryption for important data and homomorphic encryption for special data to form a hierarchical protection system. It also deploys a service modality adaptation engine to automatically switch interaction modes based on user profiles and implement cross-departmental risk prevention and control through a federal audit mechanism. The prevention and control layer integrates the data governance achievements and algorithm operation logs of the first two phases to build a risk warning matrix covering the entire technology stack. It also monitors data flow anomalies in real time and automatically triggers circuit breakers when the frequency of sensitive data access exceeds a security threshold. To address the risk of technological alienation, a public supervision interface and digital inclusion module are established. These modules automatically identify specific user groups and switch interaction modes, dynamically balancing the allocation of online and offline service resources, effectively preventing data leakage risks, improving service accessibility for vulnerable groups, and promoting the deep integration of technological empowerment and public value in intelligent government systems.

[0076] The data governance layer in this embodiment includes:

[0077] The sensitive semantic recognition engine is used to perform ternary classification of government data on certificates and government documents using a pre-trained language model (such as RoBERTa) and a graph convolutional network fusion architecture, achieving intelligent classification of government data. It also uses transfer learning to adapt to vertical domain terminology libraries to build 300+ sensitive entity recognition rules, ensuring an accuracy rate of over 98% for identifying important data. The ternary classification of government data includes important data, special data, and general data. Important data undergoes full-process manual review and quantum encryption storage; special data uses a dynamic authorization access mechanism; and general data is open for sharing after being desensitized.

[0078] The data ownership blockchain evidence storage system is used to build a government alliance chain based on Hyperledger Fabric, realize the full process of data collection, labeling, and sharing, and store it on the chain. It also develops a smart contract to automatically execute the data contribution measurement model, establishes a dynamic mapping mechanism of "data contribution-equity distribution", and eliminates barriers to data sharing between departments.

[0079] The federated learning system is used to deploy a horizontal federated learning system. Each government node retains local data, achieves cross-departmental feature fusion through encrypted parameter exchange, and adopts differential privacy technology to ensure that data is "available but invisible", breaking the dilemma of data silos while preventing privacy leakage risks.

[0080] The algorithm transparency collaboration layer in this embodiment includes:

[0081] An interpretable neural network architecture unit is used to build a GNN-based causal inference model, automatically generate a three-dimensional map of the decision path, and intuitively display the basis, data dependencies, and weight ratios of the approval decision through Attention weight visualization technology, thereby increasing the algorithm traceability efficiency by three times.

[0082] The Generative Adversarial Network Sandbox Testing Unit deploys an algorithm stress testing environment, uses GAN to generate extreme case simulation systems, and conducts adversarial attack tests on government approval scenarios. This tests the model's robustness against data poisoning and prompt injection attacks, discovering and fixing 23% of potential decision-making vulnerabilities.

[0083] The reinforcement learning parameter optimization unit is used to design a multi-objective reward function, integrate service efficiency and fairness indicators (such as the response time of digitally disadvantaged groups), and dynamically adjust algorithm parameters through the DQN algorithm to achieve an automatic balance between technology empowerment and risk prevention and control.

[0084] The prevention and control layer in this embodiment includes:

[0085] The time-series graph neural network analysis unit integrates data access logs and decision trajectory data to build a spatiotemporal feature analysis model. It also detects abnormal flow patterns of sensitive data (such as high-frequency cross-domain access) in real time, completes risk classification and triggers the circuit breaker mechanism within 200ms, significantly reducing the risk of data leakage.

[0086] A differential privacy dynamic desensitization unit is used to deploy an attribute-based encryption (ABE) gateway based on data classification labels. Important data is encrypted using quantum key distribution, special data is processed using homomorphic encryption, and general data is set with a k-anonymization threshold, achieving a dynamic balance between hierarchical protection and business efficiency.

[0087] The service modality adaptation engine is used to build a user-profile-driven intelligent dispatching system. When an elderly user is detected, the system automatically switches the voice interaction channel and increases the font contrast. When the regional network failure rate exceeds the threshold, the intelligent dispatch offline service outlets initiate an emergency response to ensure that the service coverage rate remains 100%.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A generative artificial intelligence-based e-government service method, characterized in that: The method is as follows: Build a standardized data base: Use a sensitive semantic recognition engine to extract multi-dimensional features from certificates and government documents, and combine it with a blockchain-based data ownership system to achieve full data traceability, forming a standardized and traceable government data asset pool. Establish a trusted decision-making center: Build an interpretable neural network architecture based on a standardized data base, dynamically analyze the basis and data dependencies in approval decisions through causal inference graphs, and introduce adversarial generative network sandbox testing to stress-test the algorithm, detecting model robustness in data poisoning and prompt injection risk scenarios, thereby obtaining decision traces and operation logs; Build a security assurance system: Based on decision-making trajectories and operation logs, a spatiotemporal feature analysis model is constructed to monitor abnormal data flows in real time. Differential privacy dynamic desensitization technology is used to implement quantum encryption for important data, and homomorphic encryption is enabled for special data to form a hierarchical protection system. At the same time, a service modality adaptation engine is deployed to automatically switch interaction modes based on user portraits, and cross-departmental risk joint prevention and control is achieved through a federal audit mechanism.

2. The generative artificial intelligence-based e-government service method according to claim 1 is characterized in that: The standardized data base has the following: The sensitive semantic recognition engine uses a pre-trained language model and a graph convolutional network fusion architecture to perform ternary classification of government data on certificates and government documents, enabling intelligent classification of government data. It also uses transfer learning to adapt to vertical domain terminology libraries to build over 300 sensitive entity recognition rules, ensuring an accuracy rate of over 98% for identifying important data. Dynamic mapping of cross-departmental data contribution and equity distribution: Implement intelligent measurement and equity distribution of each department's data contribution through the data ownership blockchain notarization system; Multi-source heterogeneous data fusion based on federated learning: Deploy a horizontal federated learning system, where each government node retains local data, achieves cross-departmental feature fusion through encrypted parameter exchange, and uses differential privacy technology to ensure that data is "available but invisible", solving the dilemma of data silos while preventing privacy leaks.

3. The generative artificial intelligence-based e-government service method according to claim 2 is characterized in that: The three-category classification of government data includes important data, special data and general data; among them, important data is subject to full-process manual review and quantum encryption storage; special data adopts a dynamic authorization access mechanism; general data is open for sharing after desensitizing processing.

4. The generative artificial intelligence-based e-government service method according to claim 2, characterized in that: The data ownership blockchain notarization system specifically includes: building a government alliance chain based on Hyperledger Fabric, realizing the on-chain notarization of the entire process of data collection, labeling, and sharing, and developing a smart contract to automatically execute a data contribution measurement model, establishing a "data contribution-equity distribution" dynamic mapping mechanism, and eliminating barriers to data sharing between departments.

5. The generative artificial intelligence-based e-government service method according to claim 1 is characterized in that: The details of establishing a trusted decision-making center are as follows: Interpretable neural network architecture: Builds a GNN-based causal inference model, automatically generates a three-dimensional map of the decision path, and uses attention weight visualization technology to intuitively display the basis, data dependencies, and weight ratios of approval decisions, thereby increasing algorithm tracing efficiency by three times. Adversarial Generative Network Sandbox Testing: We deployed an algorithm stress testing environment, used GAN to generate extreme case simulation systems, and conducted adversarial attack tests on government approval scenarios. We tested the model's robustness against data poisoning and prompt injection attacks, discovering and fixing 23% of potential decision-making vulnerabilities. Reinforcement learning parameter optimization: Design a multi-objective reward function that integrates service efficiency and fairness indicators, and dynamically adjust algorithm parameters through the DQN algorithm to achieve an automatic balance between technology empowerment and risk prevention and control.

6. The generative artificial intelligence-based e-government service method according to claim 1, characterized in that: The details of building a security system are as follows: Time-series graph neural network analysis: This integrates data access logs and decision trajectory data to build a spatiotemporal feature analysis model, and detects abnormal flow patterns of sensitive data in real time. It completes risk assessment and triggers circuit breakers within 200ms, significantly reducing the risk of data leakage. Differential privacy dynamic desensitization: Deploy an attribute-based encryption gateway based on data classification labels. Important data is encrypted using quantum key distribution, special data is processed using homomorphic encryption, and general data is set with a k-anonymization threshold, achieving a dynamic balance between hierarchical protection and business efficiency. Service modality adaptive engine: Builds a user-profile-driven intelligent dispatching system; automatically switches the voice interaction channel and increases font contrast when an elderly user is detected; and intelligently dispatches offline service outlets to initiate emergency responses when the regional network failure rate exceeds a threshold, ensuring 100% service coverage.

7. A generative artificial intelligence-based e-government service device, characterized in that: The device includes: The data governance layer is used to extract multi-dimensional features from certificates and government documents through a sensitive semantic recognition engine, and to achieve full data traceability through a blockchain-based data ownership system, thereby forming a standardized and traceable government data asset pool. The algorithm transparency collaboration layer is used to build an interpretable neural network architecture based on a standardized data base. It dynamically analyzes the basis and data dependencies in approval decisions through causal inference graphs. It also introduces adversarial generative network sandbox testing to stress-test the algorithm, detect model robustness under data poisoning and prompt injection risk scenarios, and obtain decision trajectories and operation logs. The prevention and control layer is used to build a spatiotemporal feature analysis model based on decision trajectories and operation logs to monitor abnormal data flows in real time, and adopt differential privacy dynamic desensitization technology to implement quantum encryption for important data, and enable homomorphic encryption for special data to form a hierarchical protection system; at the same time, a service modality adaptation engine is deployed to automatically switch the interaction mode according to the user portrait, and realize cross-departmental risk joint prevention and control through the federal audit mechanism.

8. The generative artificial intelligence-based e-government service device according to claim 7, characterized in that: The data governance layer includes: The sensitive semantic recognition engine is used to perform ternary classification of government data on certificates and government documents using a fusion architecture of pre-trained language models and graph convolutional networks, achieving intelligent classification of government data. It also uses transfer learning to adapt to vertical domain terminology libraries to build 300+ sensitive entity recognition rules, ensuring an accuracy rate of over 98% for identifying important data. The ternary classification of government data includes important data, special data, and general data. Important data undergoes full-process manual review and quantum encryption storage; special data uses a dynamic authorization access mechanism; and general data is open for sharing after being desensitized. The data ownership blockchain evidence storage system is used to build a government alliance chain based on Hyperledger Fabric, realize the full process of data collection, labeling, and sharing, and store it on the chain. It also develops a smart contract to automatically execute the data contribution measurement model, establishes a dynamic mapping mechanism of "data contribution-equity distribution", and eliminates barriers to data sharing between departments. The federated learning system is used to deploy a horizontal federated learning system. Each government node retains local data, achieves cross-departmental feature fusion through encrypted parameter exchange, and adopts differential privacy technology to ensure that data is "available but invisible", breaking the dilemma of data silos while preventing privacy leakage risks.

9. The generative artificial intelligence-based e-government service device according to claim 7, characterized in that: The algorithm transparency collaboration layer includes: An interpretable neural network architecture unit is used to build a GNN-based causal inference model, automatically generate a three-dimensional map of the decision path, and intuitively display the basis, data dependencies, and weight ratios of the approval decision through Attention weight visualization technology, thereby increasing the algorithm traceability efficiency by three times. The Generative Adversarial Network Sandbox Testing Unit deploys an algorithm stress testing environment, uses GAN to generate extreme case simulation systems, and conducts adversarial attack tests on government approval scenarios. This tests the model's robustness against data poisoning and prompt injection attacks, discovering and fixing 23% of potential decision-making vulnerabilities. The reinforcement learning parameter optimization unit is used to design a multi-objective reward function, integrate service efficiency and fairness indicators, and dynamically adjust algorithm parameters through the DQN algorithm to achieve an automatic balance between technology empowerment and risk prevention and control.

10. The generative artificial intelligence-based e-government service device according to any one of claims 7 to 9, characterized in that: The prevention and control layer includes: The time-series graph neural network analysis unit integrates data access logs and decision trajectory data to build a spatiotemporal feature analysis model and detect abnormal flow patterns of sensitive data in real time. It completes risk classification and triggers the circuit breaker mechanism within 200ms, significantly reducing the risk of data leakage. A differential privacy dynamic desensitization unit is used to deploy an attribute-based encryption gateway based on data classification labels. Important data is encrypted using quantum key distribution, special data is processed using homomorphic encryption, and general data is set with a k-anonymization threshold, achieving a dynamic balance between hierarchical protection and business efficiency. The service modality adaptation engine is used to build a user-profile-driven intelligent dispatching system. When an elderly user is detected, the system automatically switches the voice interaction channel and increases the font contrast. When the regional network failure rate exceeds the threshold, the intelligent dispatch offline service outlets initiate an emergency response to ensure that the service coverage rate remains 100%.

Images