Blockchain-based intelligent customer service AI model privacy protection system

Abstract

The application discloses a blockchain-based intelligent customer service AI model privacy protection system, belonging to the cross field of artificial intelligence and blockchain technology, and comprising five modules of model security deployment, dynamic privacy processing, blockchain permission management, smart contract auditing and abnormal monitoring optimization. The system is supported by a hybrid chain distributed ledger of a consortium chain and a public chain as a core storage support, and is combined with privacy protection algorithms of asymmetric encryption, hash encryption and zk-SNARKs zero-knowledge proof, and combined with technical means of AI model optimization, dynamic encryption adaptation, AI intent recognition and data desensitization to construct a multi-level integrated privacy protection system. The system realizes the whole life cycle privacy protection of customer service interaction data, model training data and AI models from collection, training, deployment, calling to circulation and destruction, and through dynamic encryption strategies, fine permission management, automatic compliance auditing and full-dimensional abnormal monitoring, the system gives consideration to privacy protection safety, system operation efficiency and compliance.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and blockchain technology. Specifically, it relates to a blockchain-based intelligent customer service AI model privacy protection system, which realizes privacy protection for customer service interaction data, model training data, and the AI ​​model itself throughout its entire lifecycle. Background Technology

[0002] With the deepening of digital transformation, intelligent customer service systems have been widely applied in various fields. Relying on the capabilities of artificial intelligence models, they improve service efficiency, reduce operating costs, and achieve uninterrupted service, becoming a core component of enterprise digital service systems. As intelligent customer service AI models continue to iterate, their performance improvement depends on training with large-scale user interaction data. This data includes ordinary dialogue information and a large amount of sensitive personal information, which is the core foundation for model training. This also makes privacy protection a core bottleneck for the sustainable development of intelligent customer service systems.

[0003] The privacy protection of current intelligent customer service AI models faces multiple challenges. Traditional centralized privacy protection solutions are no longer adequate for the needs of existing technological scenarios, with core pain points manifesting in three areas. First, in the data collection and storage phase, most intelligent customer service systems use centralized databases to store relevant data, resulting in issues such as plaintext data storage and chaotic access control, which easily leads to data breaches. Furthermore, some systems excessively collect user information and fail to clearly inform users of the scope of data usage, violating core principles of relevant regulations. Second, in the model training and inference phase, privacy breach risks are diversified. Model overfitting can easily lead to member inference attacks, where attackers can analyze model predictions to determine sample attribution and thus identify users. Model extraction attacks allow attackers to reverse engineer model parameters and functions by calling customer service interfaces, leaking core model technologies and indirectly disclosing sensitive information in training data. Finally, in the data sharing and transfer phase, intelligent customer service systems often need to share dialogue data with third parties for model optimization. Some third parties engage in data misuse and resale, exacerbating the risk of privacy breaches.

[0004] Currently, the application of blockchain technology in intelligent customer service scenarios still faces many compatibility issues. Existing blockchain technology has a low degree of integration with intelligent customer service AI models, lacks a targeted privacy protection architecture design, and struggles to meet the high-concurrency, low-latency service requirements of intelligent customer service. Differences in the computing power of blockchain nodes may lead to decreased model training efficiency, and node management and access control mechanisms have not yet been deeply integrated with the business scenarios of intelligent customer service, easily resulting in problems such as permission confusion and uncontrolled access. Furthermore, the development threshold for blockchain privacy protection technology is high, and its synergistic optimization with existing AI privacy protection technologies is insufficient, failing to form a mature, integrated privacy protection solution. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a blockchain-based intelligent customer service AI model privacy protection system. Through the deep integration of blockchain and privacy computing technologies, this invention achieves end-to-end privacy protection for the intelligent customer service AI model from training, deployment, and invocation to data flow, balancing privacy security, service continuity, and compliance, and solving core problems such as data leakage, abuse of permissions, and difficulty in tracing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a blockchain-based... The intelligent customer service AI model privacy protection system includes a model security deployment module, a dynamic privacy processing module, a blockchain permission control module, a smart contract auditing module, and an anomaly monitoring and optimization module. These modules work collaboratively to construct a multi-layered privacy protection system. It combines blockchain distributed ledger, asymmetric encryption, hash encryption algorithms, and zero-knowledge proof privacy enhancement algorithms. The system also incorporates blockchain's decentralized architecture, dynamic encryption strategies for AI model optimization algorithms, and adaptive AI intent recognition and data anonymization algorithms, resulting in a mature and integrated privacy protection solution.

[0007] The module includes a model security deployment module, a dynamic privacy processing module, a blockchain permission control module, a smart contract audit module, and an anomaly monitoring and optimization module. These modules integrate and adapt to various algorithm technologies and combine multiple blockchain and AI-related technologies to achieve integrated privacy protection.

[0008] This invention employs a hybrid blockchain distributed ledger as the core support for system data storage. A distributed ledger is a decentralized database technology where data is stored across multiple nodes in a network, with each node independently maintaining a complete copy of the ledger. The hybrid chain combines centralized and decentralized characteristics, flexibly adapting to different needs. Unlike traditional centralized ledgers, it does not rely on a single authoritative institution but ensures data consistency across all nodes through a consensus mechanism. It is used to record customer service interaction data, model training data, and operation records of various modules, achieving distributed management of data storage. This provides a data foundation for subsequent privacy tracing and permission verification, ensuring the integrity and accessibility of data storage.

[0009] Blockchain, with its immutability and traceability, runs through the entire system process, leaving a complete record of all operations in data collection, transmission, storage, use, and destruction. This ensures that all privacy-related operations cannot be illegally tampered with, and allows for tracing the source of operations as needed, providing support for auditing and verification.

[0010] Asymmetric encryption, hash encryption algorithms, and zero-knowledge proof privacy enhancement algorithms are the core means of protecting system privacy. Specifically, asymmetric encryption algorithms employ RSA and ECC algorithms. They work by using a pair of distinct public and private keys for encryption and decryption. The public key can be publicly used for data encryption and authentication, while the private key is exclusively kept by the system for decryption. This enables encrypted storage and transmission of data and model parameters, ensuring that data transmission and storage are not illegally intercepted. Hash encryption algorithms employ SHA-256 and MD5 algorithms. They convert input data of arbitrary length into a fixed-length hash value using a hash function. This hash value is irreversible and unique, and can be used to verify the integrity of data and model parameters, preventing data tampering. Zero-knowledge proof privacy enhancement algorithms employ zk-SNARKs algorithms, ensuring the normal training and inference of AI models without disclosing original sensitive data.

[0011] The AI ​​model optimization algorithm, adapted to the model security deployment module and anomaly monitoring optimization module, specifically employs a lightweight neural network pruning algorithm, a model quantization algorithm, and a LoRA fine-tuning algorithm. The lightweight neural network pruning algorithm reduces the model size without sacrificing core accuracy by removing redundant neurons and connections. The model quantization algorithm converts model parameters from high precision to fixed-length low precision, reducing computational complexity. The LoRA fine-tuning algorithm freezes the core parameters of the original AI model, training only the low-rank adaptation matrix, significantly reducing training computational overhead while ensuring the model's accuracy in adapting to privacy-preserving scenarios. These algorithms work synergistically to optimize the operational efficiency and performance of the intelligent customer service AI model, achieving privacy protection while avoiding any impact on model training, inference, and customer service response speed, thus ensuring a balance between system privacy protection and operational efficiency.

[0012] The dynamic encryption strategy adaptation algorithm is mainly integrated into the dynamic privacy processing module. It adopts the adaptive encryption strength adjustment algorithm (AESA), the scene-aware dynamic encryption algorithm (SADE), and the privacy level driven encryption adaptation algorithm (PLDEA). It can dynamically adjust the encryption strategy and encryption strength according to the privacy level of the data, the usage scenario, and the flow process, so as to adapt to the privacy protection needs of different types of data and improve the flexibility and pertinence of the system's privacy protection.

[0013] The AI ​​intent recognition and data anonymization algorithm, belonging to the dynamic privacy processing module, accurately distinguishes between sensitive and non-sensitive information in customer service interactions and model training data through AI intent recognition. Combined with data anonymization methods, it processes sensitive information to avoid the risk of sensitive data leakage at the source, laying the foundation for subsequent privacy protection.

[0014] Beneficial effects

[0015] (1) Enhance the security and comprehensiveness of privacy protection. Through various encryption algorithms, blockchain decentralized architecture and dynamic adaptation strategies, achieve full lifecycle privacy protection for customer service interaction data, model training data and AI models themselves, effectively avoiding security risks such as data leakage and tampering.

[0016] (2) Ensure a balance between system operating efficiency and model performance. Through new AI model optimization algorithms, while achieving comprehensive privacy protection, reduce the impact of protection measures on AI model training, inference and intelligent customer service response speed, and ensure stable and efficient system operation.

[0017] (3) Enhance the flexibility and compliance of privacy protection. Dynamic encryption strategies can adapt to the protection needs of data with different privacy levels. Combined with the traceability and immutability of blockchain and related compliance verification designs, it ensures that system operation and data processing comply with relevant privacy protection regulations. Attached Figure Description

[0018] Figure 1 This is a diagram of the overall system architecture of the present invention.

[0019] Figure 2 This is a flowchart illustrating the specific process of this invention.

[0020] Figure 3 This is a diagram of the hybrid chain blockchain distributed ledger architecture of the present invention.

[0021] Figure 4 This is a structural diagram of the privacy enhancement algorithm combining asymmetric encryption, hash encryption, and zero-knowledge proof in this invention.

[0022] Figure 5 This diagram illustrates the combination of the model quantization algorithm and the LoRA fine-tuning algorithm in this invention. Detailed Implementation

[0023] This invention relates to a blockchain-based intelligent customer service AI model privacy protection system. The system relies on a hybrid blockchain distributed ledger as its core storage support, integrating privacy protection algorithms such as asymmetric encryption, hash encryption, and zero-knowledge proofs. It combines AI model optimization, dynamic encryption adaptation, intent recognition, and data anonymization techniques. Through the collaborative operation of a model security deployment module, a dynamic privacy processing module, a blockchain permission control module, a smart contract audit module, and an anomaly monitoring and optimization module, it achieves full lifecycle privacy protection for the intelligent customer service AI model, from data collection, model training, deployment, and data flow to destruction. The following details the specific implementation process, technical implementation methods, and overall collaborative operation mechanism of each module. This implementation method is only for explaining the invention and is not intended to limit the scope of protection of the invention.

[0024] I. System Deployment and Basic Configuration 1. The hybrid blockchain distributed ledger is built using a hybrid blockchain architecture combining a consortium blockchain and a public blockchain. The consortium blockchain nodes consist of the company owning the intelligent customer service system, a compliance auditing agency, and authorized model optimization partners. These nodes are responsible for storing sensitive information such as core customer service interaction data, model training data, and system operation records. A PBFT (Practical Byzantine Fault Tolerance) consensus mechanism is used to ensure data consistency and transaction processing efficiency, adapting to the high concurrency and low latency service requirements of intelligent customer service. The public blockchain nodes store de-identified non-sensitive data, data flow traceability identifiers, and other public information. A PoS (Proof-of-Stake) consensus mechanism is used to reduce computational power consumption. The ledger's underlying layer deploys data sharding storage technology, distributing data of different privacy levels across corresponding chain nodes. Each data unit is assigned a unique hash identifier, recording metadata such as data collection time, source, and processing stage, providing a foundation for subsequent traceability and verification. After the ledger deployment is complete, initial permission configurations are performed on all nodes, binding node identities and operation permissions to ensure that only authorized nodes can access the corresponding data. 2. Corresponding core algorithms are pre-integrated into each module of the system. Specifically, the model security deployment module integrates lightweight neural network pruning, model quantization, and LoRA fine-tuning algorithms; the dynamic privacy processing module integrates AESA adaptive encryption strength adjustment, SADE scene-aware dynamic encryption, PLDEA privacy-level driven encryption adaptation, and AI intent recognition and data anonymization algorithms; the blockchain permission control module integrates RSA and ECC asymmetric encryption algorithms, and SHA-256 and MD5 hash encryption algorithms; and the entire system uniformly integrates zk-SNARKs zero-knowledge proof privacy enhancement algorithms. Simultaneously, an algorithm call interface and parameter adjustment platform are built to support dynamic adjustment of algorithm parameters according to business needs. 3. Hardware and software environment adaptation: Matching computing power hardware is configured for blockchain nodes. For core nodes of consortium blockchains, computing power configuration is improved to ensure model training and data processing efficiency; for public blockchain nodes, lightweight computing power configuration is adopted to reduce operation and maintenance costs. On the software side, an integrated operating system compatible with blockchain and AI model operation is built, the system interaction interface is optimized to ensure smooth data transmission and command calls between modules, and it is also connected to the existing intelligent customer service system to achieve seamless integration without affecting the original customer service process.

[0025] II. Specific Implementation Process of Each Module (I) Model Security Deployment Module: This module enables secure and lightweight deployment and training of AI models. As the foundational support for the system's models, it focuses on lightweight optimization, privacy-preserving training, and secure deployment of the intelligent customer service AI model. This prevents the model from negatively impacting privacy and service experience due to parameter leaks or low training efficiency. The specific implementation steps are as follows: First, lightweight optimization of the original model involves using a lightweight neural network pruning algorithm on the existing intelligent customer service AI model. Through model training effect evaluation, redundant neurons and connections with contributions below a preset threshold are removed. After removal, the model's accuracy is verified. If the accuracy loss exceeds 5%, the pruning threshold is adjusted retrospectively to ensure a 30%-50% reduction in model size without sacrificing core service accuracy. Next, a model quantization algorithm is used to convert the 32-bit floating-point parameters in the model into 8-bit integer parameters, reducing computational complexity and improving inference and response speed. Privacy-preserving fine-tuning training is then performed based on the optimized lightweight model, using the LoRA fine-tuning algorithm. The core parameters of the model are frozen, and only a low-rank adaptation matrix is ​​constructed for fine-tuning according to the intelligent customer service business scenario and privacy protection requirements. Training data uses encrypted data that has been desensitized by a dynamic privacy processing module. During training, the zk-SNARKs zero-knowledge proof algorithm is used to update model parameters without revealing the original training data information. Simultaneously, parameter adjustment records and data usage records are stored on the hybrid consortium blockchain node in real time, ensuring traceability of the training process. For secure model deployment and on-chain notarization, the fine-tuned intelligent customer service AI model is encrypted using the RSA asymmetric encryption algorithm to encrypt core model parameters. The private key is exclusively stored by the system, while the public key is used for authentication after model deployment. Model deployment employs a distributed deployment approach, splitting the model across multiple core nodes on the consortium blockchain to prevent model failure due to a single node malfunction. Key data such as the model's deployment location, encryption information, and version number are stored on the blockchain as a unique identifier for each deployment. Subsequent model updates and invocations require authorization verification through this identifier.

[0026] (II) Dynamic Privacy Processing Module: Achieving Dynamic Privacy Protection Throughout the Data Processing Process. This module is the core privacy processing unit of the system, responsible for identifying and processing sensitive data at the source. It dynamically adjusts encryption strategies based on data privacy levels and usage scenarios to adapt to the privacy protection needs of different stages. The specific implementation steps are as follows: 1. AI Intent Recognition and Sensitive Data Differentiation: When the intelligent customer service system generates interaction data and model training data, the data first enters this module. The AI ​​intent recognition algorithm performs semantic analysis and feature extraction on the data. A sensitive information recognition database is preset, and the algorithm matches the data with the database to accurately distinguish between sensitive and non-sensitive data, and labels the sensitive data with a privacy level. 2. Tiered Data Desensitization: The differentiated sensitive data undergoes tiered desensitization. Irreversible desensitization is used for high-privacy-level data; reversible desensitization is used for medium-privacy-level data; and obfuscation is used for low-privacy-level data. Non-sensitive data is directly hashed to generate a unique hash value, completing the initial data processing. The processed data metadata is recorded synchronously. 3. Dynamic Encryption Strategy Adaptation: The de-identified dataset undergoes encryption processing based on usage scenarios and workflow stages through dynamic encryption strategy adaptation algorithms. The PLDEA privacy-level driven encryption adaptation algorithm matches the basic encryption algorithm to the data's privacy level. High-privacy-level data uses dual encryption with ECC asymmetric encryption and zk-SNARKs zero-knowledge proofs; medium-privacy-level data uses RSA asymmetric encryption; and low-privacy-level data uses SHA-256 hash encryption. The SADE scene-aware dynamic encryption algorithm reduces encryption computation complexity and ensures response speed for real-time customer service interactions; it increases encryption strength to prevent data leakage during model training; and it uses a dedicated encryption key for third-party data sharing, allowing only authorized third parties to decrypt the data. The AESA adaptive encryption strength adjustment algorithm monitors the network environment and computing load for data transmission and usage in real time. If a security risk is detected in the network environment, the encryption strength is automatically increased; if the computing load is too high, the encryption strength is appropriately reduced within a safe threshold to balance privacy protection and system efficiency. 4. Encrypted Data Transmission and Storage: The encrypted dataset is sent to the hybrid chain distributed ledger through an encrypted transmission channel. Depending on the privacy level and usage requirements, it is stored on either a consortium chain or a public chain node. During storage, a hash verification is performed again to ensure that the data transmission has not been tampered with. After the data storage is completed, a data storage address and traceability identifier are generated and synchronized to the blockchain permission control module.(III) Blockchain Access Control Module: Achieving Refined Access Management and Data Access Control Across the Entire System. This module leverages the decentralized and immutable characteristics of blockchain to achieve access control over all system nodes, operators, and partners, as well as precise authorization for data access and model calls, preventing abuse of permissions and unauthorized access. The specific implementation steps are as follows: 1. Identity Authentication and Access Control Hierarchy: Establish a unified identity authentication system, assigning unique digital identities to system operators, blockchain nodes, third-party partners, and intelligent customer service terminals. RSA asymmetric encryption algorithm is used for identity signing and verification. Digital identities are bound to the entity's real information and business permissions. All identity information is stored on the blockchain node and cannot be tampered with. Access is hierarchically classified according to the "principle of least privilege," divided into five levels: super administrator privileges, operation privileges, partner privileges, and terminal privileges. Each level is configured with corresponding operation scope and data access permissions. 2. Data Access and Model Call Authorization: When an entity requests data access or model call, it must submit authentication information and request content to this module. The module first verifies the identity information. After successful verification, it determines whether the entity has the corresponding operation permission based on its permission level. If the entity has permission, a one-time access key is automatically generated. The key's validity period is set according to the purpose of the request, and the access request, authorization result, and key generation information are recorded on the blockchain. If the entity does not have permission, the request is directly rejected, and the abnormal request information is synchronized to the anomaly monitoring and optimization module. After the entity completes data access or model call using the one-time access key, the key automatically expires, ensuring that data and models are only used by authorized entities within the specified scope. 3. Dynamic adjustment and revocation of permissions. Supports dynamic adjustment of subject permissions according to business needs. Super administrators can submit permission adjustment applications through the module. The application must include the reasons and basis for the adjustment. The adjustment information will take effect after being recorded on the chain. If abnormal operation of a subject is detected, the module will automatically freeze its permissions temporarily and synchronize them to the abnormal monitoring and optimization module. After verification, it will decide whether to restore or permanently revoke the permissions. After the cooperation period of a third-party partner expires, the module will automatically revoke all its permissions and destroy the corresponding access key to ensure the timeliness of permission control. (IV) Smart contract audit module: realize automated audit and compliance verification of the entire system operation process. This module is based on blockchain smart contract technology and pre-writes audit and compliance verification contracts to realize automated and real-time audit of all operations such as system data collection, transmission, storage, use, destruction, model training, deployment, and calling. It ensures that the system operation complies with privacy protection regulations and internal enterprise systems. The specific implementation steps are as follows: 1. Smart contract writing and deployment. Combining privacy protection regulations and enterprise business needs, multiple smart contracts are written, including data processing compliance contracts, operation audit contracts, and model management contracts. The completed smart contract undergoes security testing to avoid vulnerabilities. Once the test is passed, it is deployed to the hybrid chain consortium chain node. Once deployed, the contract cannot be tampered with, and all system operations must trigger the corresponding contract for verification.2. Full-Process Operation Triggering and Contract Verification: All operations in each module of the system must trigger smart contracts in real time for compliance verification. During data collection, the data processing compliance contract is triggered to verify whether explicit user authorization has been obtained and whether excessive data collection exists. If the verification passes, collection is allowed; otherwise, collection is prohibited and the reason is displayed. During model training, the model management contract is triggered to verify whether the training data has been anonymized and encrypted and whether authorized data is used. If the verification passes, training is allowed; otherwise, training is terminated and an anomaly is recorded. During third-party data sharing, the data processing compliance contract is triggered to verify whether the shared data has been anonymized and whether the third party has the corresponding permissions. If the verification passes, data transfer is allowed; otherwise, sharing is prohibited. All operation triggering contract processes and verification results are stored on the blockchain in real time, forming an immutable audit record. 3. Automatic Audit Report Generation and Compliance Early Warning Module: Based on the audit records stored on the blockchain, the module automatically generates system audit reports according to a preset cycle. The reports include operation statistics, compliance verification results, and details of abnormal operations. After generation, the reports are synchronized to the super administrator and audit institution nodes. If a compliance risk is detected during contract verification, the module will immediately trigger a compliance warning, notifying the super administrator and auditors via the system backend, SMS, etc., while simultaneously locking related operations to prevent the risk from escalating. The warning information and processing process are recorded on the blockchain throughout. 4. Smart Contract Iteration and Optimization: When privacy protection regulations are updated or business needs are adjusted, the super administrator can submit a contract iteration request. In conjunction with the auditing firm, the original contract will be modified and optimized. The optimized contract must undergo security testing and on-chain notarization before replacing the original contract and taking effect. The entire iteration process is traceable. (V) Anomaly Monitoring and Optimization Module: Realizing System Anomaly Detection, Risk Handling, and Continuous Model Optimization. This module serves as the monitoring and optimization hub of the system, monitoring the operational status of each module, data flow process, and model performance in real time. It promptly detects and handles various anomalies and risks, and continuously optimizes the AI ​​model and privacy protection strategies based on monitoring data to ensure stable and efficient system operation. The specific implementation steps are as follows: 1. Multi-dimensional Real-time Monitoring: Establishing a multi-dimensional monitoring system to monitor the system 24 / 7. The monitoring scope includes: System operational status: blockchain node computing power, load, network connection status, data transmission speed of each module, and algorithm call efficiency; Data flow process: risks such as abnormal access, data tampering, and key leakage in the data collection, transmission, storage, and usage stages; Model performance: response speed, answer accuracy, and parameter leakage risks in the inference process of the intelligent customer service AI model; Permission operation behavior: behaviors such as unauthorized access by the subject, multiple failed authentication attempts, and abnormal key usage. Normal thresholds are set for each monitoring indicator. If an indicator exceeds the threshold, it is judged as an anomaly, triggering the anomaly detection mechanism.2. Anomaly Risk Classification and Handling: Detected anomalies are classified into low-level, medium-level, and high-level anomalies, with automated handling strategies developed for each level: Low-level anomalies: The system automatically adjusts parameters and sends an alert to operations and maintenance personnel, who promptly verify and handle the issue. High-level anomalies: The system immediately triggers an emergency mechanism, severing the connection between the abnormal node and the system, destroying leaked keys, suspending external model services, and sending an emergency alert to the super administrator and security team. Manual risk handling is then organized, and the entire process is recorded on the blockchain. After anomaly handling is completed, the cause of the anomaly is analyzed, and corrective measures are developed to prevent similar anomalies from recurring. 3. The AI ​​model continuous optimization module periodically optimizes the model based on the monitoring and collection of intelligent customer service AI model operation data and training data feedback: If the model's response accuracy decreases, the LoRA fine-tuning algorithm is used to make minor adjustments to the model based on the latest desensitized training data, freezing the backbone parameters and only updating the low-rank adaptation matrix to reduce training costs; if the model's response speed decreases, a lightweight neural network pruning algorithm and model quantization algorithm are used again to optimize the model a second time, further reducing the model size and computational complexity; in combination with new user interaction scenarios and new needs, the model's intent recognition library is updated to improve the model's adaptability to new scenarios. After the model optimization is completed, accuracy and performance verification are performed. After successful verification, the data is stored on the blockchain and the model is updated. 4. Privacy Protection Strategy Iteration: Based on abnormal risk records, compliance verification results, and changes in business needs during system operation, the privacy protection strategy is iteratively optimized: If a security vulnerability is detected in the encryption algorithm of a certain type of privacy-level data, it is immediately replaced with a more secure encryption algorithm, and the stored data of that type is re-encrypted; if it is found that the dynamic encryption strategy for a certain business scenario cannot balance privacy protection and operational efficiency, the parameters of AESA, SADE, and PLDEA algorithms are adjusted, and the encryption strategy adaptation logic is optimized; in accordance with new privacy protection regulations, the sensitive information identification database and data anonymization rules are updated to improve system compliance. After the privacy protection strategy is iterated, a small-scale test is conducted within the system. After the test is passed, it is fully rolled out, and the iteration information is recorded on the blockchain.

[0027] III. System-wide lifecycle collaborative operation mechanism The privacy protection system for the intelligent customer service AI model of this invention does not operate independently, but rather forms a collaborative operation mechanism with data interoperability, command linkage, and full traceability, covering the entire lifecycle of intelligent customer service from data collection to model destruction. The specific collaborative operation process is as follows: 1. Data collection stage: When the user interacts with the intelligent customer service terminal, the data is uploaded to the dynamic privacy processing module via the terminal to complete intent recognition, sensitive data differentiation and desensitization processing, and then encryption is completed through dynamic encryption strategy adaptation. The encrypted data is synchronized to the blockchain permission control module for identity verification. After the verification is passed, it is stored in the hybrid chain distributed ledger. At the same time, the data collection operation triggers the compliance verification of the smart contract audit module, and the verification result and data storage information are recorded on the chain. 2. Model training stage: The model security deployment module retrieves the encrypted and desensitized training data from the hybrid chain ledger and completes privacy training through the zk-SNARKs zero-knowledge proof algorithm. During the training process, the model optimization algorithm collaboratively improves the training efficiency. The training operation triggers the smart contract audit module for verification. The training record and model parameter update information are stored on the chain for evidence. The anomaly monitoring and optimization module monitors the training process in real time and detects the risk of computing power load and model parameter leakage. 3. Model Deployment and Customer Service Phase: The model security deployment module completes the encryption and distributed deployment of the optimized model, and stores the deployment information on the blockchain. The intelligent customer service terminal submits a model call request, which is authenticated and authorized by the blockchain permission control module, generating a one-time access key. The terminal uses the key to call the model for real-time customer service interaction. The dynamic privacy processing module performs real-time desensitization and encryption on new data generated during the interaction. The anomaly monitoring and optimization module monitors the model's response speed, accuracy, and system operating status to ensure a good service experience. 4. Data Sharing and Transfer Phase: When a third-party partner submits a data sharing request, after authorization by the blockchain permission control module, the dynamic privacy processing module performs secondary desensitization and proprietary encryption on the shared data. The data transfer process triggers compliance verification by the smart contract audit module, and the transfer record is stored on the blockchain. After data sharing, the anomaly monitoring and optimization module monitors the third party's data usage behavior to prevent data abuse. 5. Data and Model Destruction Phase: When data reaches its storage expiration date or the model completes its lifecycle, the super administrator submits a destruction request. After verification by the blockchain permission control module, the system irreversibly destroys the data and encrypts and destroys the model's core parameters. The destruction operation triggers verification by the smart contract audit module, and the destruction record is stored on the blockchain to ensure that the data and model cannot be illegally recovered. 6. Full-Process Monitoring and Optimization: The anomaly monitoring and optimization module monitors the operational status of all the above stages in real time, promptly detects and handles anomalies, and provides optimization suggestions to each module based on the monitoring data. The model security deployment module and dynamic privacy processing module iteratively optimize the model and privacy protection strategies based on the suggestions. The blockchain permission control module and smart contract audit module update permissions and contracts according to business changes, achieving continuous system optimization.

[0028] IV. System Testing and Maintenance 1. After system deployment, conduct multi-dimensional testing, including privacy protection testing, performance testing, compliance testing, and stability testing. Record any issues and vulnerabilities identified during testing, rectify them, and retest until all test indicators meet preset standards. 2. Establish a daily system operation and maintenance (O&M) system. Assign professional O&M personnel to regularly inspect the blockchain nodes and the operational status of each module, promptly addressing hardware failures and software vulnerabilities; regularly back up system data to prevent data loss; regularly update encryption algorithms and security tools to improve the system's resistance to attacks; provide professional training to O&M personnel to ensure they are proficient in system operation and anomaly handling procedures. All operations during O&M must be recorded on the blockchain to achieve traceability.

[0029] This implementation method, through the aforementioned technical means, achieves deep integration of blockchain and privacy computing technology in the privacy protection scenario of intelligent customer service AI model. It solves the problems of data leakage, abuse of permissions, and difficulty in tracing existing in traditional privacy protection solutions, while taking into account both privacy protection security and system operating efficiency, ensuring that the intelligent customer service system operates stably and efficiently under the premise of compliance.

Claims

1. A privacy protection system for a blockchain-based intelligent customer service AI model, characterized in that, The system includes a model security deployment module, a dynamic privacy processing module, a blockchain permission control module, a smart contract auditing module, and an anomaly monitoring and optimization module. The modules are interconnected, their commands are linked, and the entire process is traceable. The system uses a hybrid blockchain distributed ledger as its core storage support, integrating asymmetric encryption algorithms, hash encryption algorithms, and zero-knowledge proof privacy enhancement algorithms. Combined with AI model optimization algorithms, dynamic encryption strategy adaptation algorithms, and AI intent recognition and data desensitization algorithms, the system achieves full lifecycle privacy protection for customer service interaction data, model training data, and the AI ​​model itself, from collection, training, deployment, and invocation to circulation and destruction. The hybrid blockchain distributed ledger adopts a hybrid blockchain architecture of consortium blockchain + public blockchain. The consortium blockchain nodes are composed of the enterprise to which the intelligent customer service system belongs, the compliance audit institution, and the authorized model optimization partners. It adopts the PBFT practical Byzantine fault-tolerant consensus mechanism to store core customer service interaction data, model training data, and system operation records. The public blockchain nodes adopt the Proof-of-Stake (PoS) consensus mechanism to store desensitized non-sensitive data and data flow traceability identifiers; the ledger's underlying layer deploys data sharding storage technology, assigning a unique hash identifier to each data unit and recording metadata, and all nodes are bound to identity and operation permissions.

2. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 1, characterized in that, The asymmetric encryption algorithms are RSA and ECC, used for encrypted storage and transmission of data and model parameters, as well as identity signing and verification; the hash encryption algorithms are SHA-256 and MD5, used to verify the integrity of data and model parameters; the zero-knowledge proof privacy enhancement algorithm is zk-SNARKs, used to ensure normal training and inference of AI models without disclosing original sensitive data.

3. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 2, characterized in that, The AI ​​model optimization algorithm includes a lightweight neural network pruning algorithm, a model quantization algorithm, and a LoRA fine-tuning algorithm, which are integrated into the model security deployment module and the anomaly monitoring and optimization module. The lightweight neural network pruning algorithm removes redundant neurons and connections in the model whose contribution is lower than a preset threshold, reducing the model size by 30%-50% without losing more than 5% of the core service accuracy. The model quantization algorithm converts the 32-bit floating-point parameters in the model into 8-bit integer parameters. The LoRA fine-tuning algorithm achieves privacy-preserving fine-tuning of the model by freezing the backbone parameters of the original AI model and training only the low-rank adaptation matrix.

4. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 3, characterized in that, The dynamic encryption strategy adaptation algorithm includes an adaptive encryption strength adjustment algorithm, a scene-aware dynamic encryption algorithm, and a privacy level-driven encryption adaptation algorithm, all integrated into the dynamic privacy processing module. The privacy level-driven encryption adaptation algorithm matches the basic encryption algorithm according to the data privacy level. High privacy level data uses dual encryption with ECC asymmetric encryption and zk-SNARKs zero-knowledge proof, medium privacy level data uses RSA asymmetric encryption, and low privacy level data uses SHA-256 hash encryption. The scene-aware dynamic encryption algorithm adjusts the encryption strategy according to scenarios such as real-time customer service interaction, model training, and third-party data sharing. The adaptive encryption strength adjustment algorithm dynamically adjusts the encryption strength based on network environment security risks and computing load.

5. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 4, characterized in that, The AI ​​intent recognition and data desensitization algorithm is integrated into the dynamic privacy processing module. It distinguishes between sensitive and non-sensitive data by combining semantic analysis and feature extraction with a sensitive information identification library, and labels the privacy level of sensitive data. Irreversible desensitization is used for high privacy level data, reversible desensitization is used for medium privacy level data, and obfuscation is used for low privacy level data. Non-sensitive data is directly hashed to generate a unique hash value.

6. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 5, characterized in that, The workflow of the model security deployment module is as follows: (1) The original intelligent customer service AI model is optimized by using a lightweight neural network pruning algorithm and a model quantization algorithm; (2) Based on the optimized lightweight model, the LoRA fine-tuning algorithm is used, and the encrypted training data after desensitization is combined with the zk-SNARKs algorithm to complete the privacy fine-tuning training, and the training record is uploaded to the blockchain in real time. (3) The core parameters of the fine-tuned model are encrypted using the RSA algorithm, the model is distributed and deployed to the core nodes of the consortium chain, and the deployment location, encryption information, and version number are stored on the chain to form a unique identifier.

7. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 6, characterized in that, The blockchain permission control module establishes a unified identity authentication system, assigning unique digital identities to system operators, blockchain nodes, third-party partners, and intelligent customer service terminals, and using the RSA algorithm to complete identity verification. The digital identities are bound to business permissions and stored on the blockchain. According to the principle of least privilege, permissions are divided into levels such as super administrator permissions, operation permissions, partner permissions, and terminal permissions, and corresponding operation and data access scopes are configured. When a data access or model call request is received, the identity and permissions are verified, a one-time access key is generated for the authorization request, the request and authorization information are recorded on the blockchain, and the key automatically expires after use. It supports dynamic adjustment and revocation of permissions. Abnormal operations trigger temporary freezing of permissions, and permissions are automatically revoked and access keys are destroyed upon the expiration of third-party cooperation.

8. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 7, characterized in that, The smart contract auditing module pre-writes and deploys data processing compliance contracts, operation audit contracts, and model management contracts to the hybrid chain consortium chain nodes. The contracts are security tested and cannot be tampered with after deployment. Each module of the system triggers the corresponding contract for compliance verification in real time, and the verification results and operation process are stored on the blockchain. The modules generate audit reports according to the on-chain audit records at preset cycles and synchronize them to the super administrator and audit institution nodes. When a compliance risk is detected, an alert is triggered and related operations are locked. The alert and handling process are recorded on the blockchain. It supports iterative optimization of smart contracts by combining regulatory updates and business needs. The optimized contracts take effect after security testing and on-chain notarization.

9. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 8, characterized in that, The anomaly monitoring and optimization module establishes a 24 / 7 multi-dimensional monitoring system. The monitoring scope includes system operation status, data flow process, model operation effect, and permission operation behavior. Normal thresholds are set for monitoring indicators, and exceeding the threshold is judged as an anomaly. Anomalies are divided into low, medium, and high levels, and automated handling strategies are formulated. High-level anomalies trigger an emergency mechanism, and the anomaly handling process and cause analysis are recorded on the blockchain. Based on the monitoring data, the AI ​​model is continuously optimized using LoRA fine-tuning algorithm, lightweight neural network pruning algorithm, and model quantization algorithm. After the optimized model is verified, it is updated on the blockchain. Combining anomaly risk records, compliance verification results, and changes in regulations and business needs, the privacy protection strategy is iteratively optimized. After strategy iteration, it is tested on a small scale before being fully promoted and recorded on the blockchain.

10. The privacy protection system for a blockchain-based intelligent customer service AI model according to claim 9, characterized in that, The system modules form a full life cycle collaborative operation mechanism, specifically: (1) Data collection stage: After the interactive data is identified, desensitized and encrypted by the dynamic privacy processing module, it is stored in the hybrid chain after identity verification by the blockchain permission control module. The collection operation triggers the compliance verification of the smart contract audit module, and all information is uploaded to the chain; (2) Model training stage: The model security deployment module retrieves the encrypted and desensitized training data, and completes privacy training by combining the zk-SNARKs algorithm. The training operation triggers the compliance verification, and the anomaly monitoring and optimization module monitors the training process in real time; (3) Model deployment and customer service stage: The model is encrypted and distributed and stored on the blockchain. After the terminal is authorized, it calls the model. The dynamic privacy processing module desensitizes and encrypts the new interactive data in real time. The anomaly monitoring and optimization module monitors the model and system operation status. (4) Data sharing and circulation stage: After the third party is authorized, the dynamic privacy processing module desensitizes and encrypts the shared data a second time. The circulation operation triggers compliance verification. The anomaly monitoring and optimization module monitors the third party's data usage behavior. (5) Data and model destruction stage: After the super administrator submits the destruction application and it is verified, the system irreversibly destroys the data and encrypts and destroys the core parameters of the model. The destruction operation triggers compliance verification and is stored on the blockchain. (6) Full process monitoring and optimization: The anomaly monitoring and optimization module monitors each stage in real time and provides feedback on optimization suggestions. Each module completes the iterative update of the model, privacy protection strategy, permissions, and contract.

Images