Model encryption and privacy protection method oriented to artificial intelligence algorithm

Abstract

The invention relates to the technical field of data processing, in particular to an artificial intelligence algorithm-oriented model encryption and privacy protection method, which comprises the following steps of: generating a privacy level mapping table through quantum feature extraction and multi-objective optimization, and dynamically screening node subsets with low delay and complementary data distribution by adopting a quantum annealing algorithm, a federated aggregation formula is improved by combining a weight prediction model of Bayesian optimization, and parameter conflicts caused by non-independent identically distributed data are relieved. For the balance problem of privacy protection and model precision, incremental update parameters are generated through quantum hash comparison, violation data streams are blocked in combination with anomaly detection of a quantum support vector machine, and an auditing chain drives a quantum hybrid optimization framework to synchronously adjust encryption strength and noise rules. And the closed loop iteration mechanism feeds back optimization parameters to the data preprocessing and encryption module to form full-link adaptive cooperation from feature dimension reduction, model training to reasoning output, so that global optimization of privacy protection and model effectiveness in a data heterogeneous scene is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a method for model encryption and privacy protection for artificial intelligence algorithms. Background Art

[0002] The model encryption and privacy protection technology for artificial intelligence algorithms aims to achieve efficient training and inference of models through cryptographic means and distributed collaboration mechanisms while ensuring data privacy. Among them, federated learning, as the mainstream privacy protection framework, stores data locally at distributed nodes and collaborates to update the global model, avoiding the centralized transmission of original data and effectively reducing the risk of privacy leakage. Differential privacy, on the other hand, injects controllable noise into model parameters or outputs to prevent individual information from being reverse-inferred. Such technologies are widely used in highly sensitive fields such as finance and healthcare to meet the increasingly strict data security regulations.

[0003] Existing federated learning frameworks face significant challenges in dealing with data heterogeneity. The non-independent and identically distributed characteristics of the participating parties' data lead to a deviation between the local model update direction and the global objective, and traditional static aggregation strategies are difficult to dynamically adjust the weight allocation, resulting in a slowdown in the model convergence speed or a decrease in accuracy. For example, the aggregation method using fixed weighted average cannot adapt to the dynamic changes in the data volume, feature distribution, or contribution degree of different nodes, exacerbating the conflict of local model parameters. In addition, the balance problem between the privacy protection mechanism and model performance in a heterogeneous environment has not been effectively solved: over-reliance on noise injection or encryption layers will further weaken the convergence stability of the model, while simplifying the protection measures may expose privacy vulnerabilities, forming a double contradiction between security and practicality. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technology, the present invention provides a method for model encryption and privacy protection for artificial intelligence algorithms, which solves the double contradiction formed by the decrease in model aggregation efficiency caused by data heterogeneity in the federated learning framework and the negative impact of the privacy protection mechanism on the convergence stability and accuracy of the model, and needs to solve the problem of parameter conflict caused by non-independent and identically distributed data between distributed nodes and the lack of a dynamic privacy regulation mechanism.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: The method for model encryption and privacy protection for artificial intelligence algorithms provided by the present invention includes: Step S101, obtaining the original business data, identifying sensitive attributes through quantum feature extraction and dimensionality reduction processing, and generating a dimensionality-reduced feature vector; Step S102, inputting the dimensionality-reduced feature vector into a multi-objective optimization and hierarchical algorithm to generate a privacy level mapping table including data sensitivity levels and the cleaned structured data; Step S103: Based on the privacy level mapping table, dynamically adjust the aggregation weights of participating nodes in the federated learning framework, select a subset of low-latency nodes through quantum optimization, and generate global model parameters. Step S104: Perform quantum fully homomorphic encryption on the highly sensitive layers in the global model parameters according to the privacy level mapping table, and perform quantum sparse compression on the encrypted parameters. Step S105: Inject dynamic noise generated based on reinforcement learning into the compressed encrypted parameters to generate noisy encrypted model parameters. Step S106: Locate the incrementally updated part of the noisy encrypted model parameters through quantum hash comparison, generate an increment list and update the global model. Step S107: Receive a user inference request, use quantum anomaly detection to identify high-frequency query behaviors, and perform fuzzification processing on the output results of the updated global model. Step S108: Intercept illegal transmissions according to compliance rules matched by data geographical tags, and generate an audit chain through quantum secure hashing. Step S109: Dynamically adjust the encryption strength and noise injection strategy based on the risk indicators in the audit chain, and feedback the adjusted policy parameters to the federated learning node selection and encryption module.

[0006] Furthermore, for the model encryption and privacy protection method for artificial intelligence algorithms of the present invention, the step S101 includes: Extract the feature vectors associated with sensitive attributes by parallel computing the feature covariance matrix through quantum principal component analysis. Input the feature vectors into a multi-objective optimization grading algorithm, and simultaneously use data sensitivity grading, service availability scoring, and GDPR compliance indicators as optimization objectives to generate a Pareto optimal solution set. Generate a privacy level mapping table according to the optimal solutions in the Pareto optimal solution set. The privacy level mapping table includes encryption level marks for each data field, and is used to control the encryption strength of the local training parameters of federated learning participating nodes.

[0007] Furthermore, for the model encryption and privacy protection method for artificial intelligence algorithms of the present invention, the step S103 includes: Construct a Gaussian process model based on a Bayesian optimizer, and predict the weight values of each federated learning participating node according to the historical aggregation contribution degrees of each federated learning participating node. Map the node communication delay and data distribution similarity to the Hamiltonian of the Ising model, and solve for the optimal subset of federated nodes through a quantum annealing machine. Adjust the weighted average coefficient in the federated aggregation formula according to the predicted weight values of each federated learning participating node and the optimal federated node subset, and generate global model parameters.

[0008] Further, in the model encryption and privacy protection method for artificial intelligence algorithms according to the present invention, the step S104 includes: Encrypt the global model parameters marked as the high-sensitivity level in the privacy level mapping table by applying the quantum fully homomorphic encryption algorithm; Perform compression processing on the encrypted parameters through the sparse coding layer of the quantum neural network to generate compressed ciphertext parameters; Input the compressed ciphertext parameters into the homomorphic operation interface of the differential privacy noise injection module to maintain noise superposition in the ciphertext state.

[0009] Further, in the model encryption and privacy protection method for artificial intelligence algorithms according to the present invention, the step S105 includes: Through the deep deterministic policy gradient reinforcement learning model, use the privacy budget consumption rate and the model accuracy loss as the reward function to generate the model layer position and variance intensity policy for noise injection; Use a quantum random number generator to generate a noise sequence that conforms to the Gaussian distribution; Through the homomorphic addition operation of quantum fully homomorphic encryption, superimpose the noise sequence on the ciphertext of the encrypted parameters to generate encrypted model parameters with noise.

[0010] Further, in the model encryption and privacy protection method for artificial intelligence algorithms according to the present invention, the step S106 includes: Based on the Grover algorithm, perform a quantum state comparison on the current global model parameters and the historical version, locate the newly added or modified parameters, and generate an incremental parameter list; Perform quantum sparse coding processing on the incremental parameter list, and generate compressed ciphertext through the quantum fully homomorphic encryption algorithm; Input the compressed ciphertext into the blockchain storage interface of the audit tracking module, trigger the global model update, and generate updated global model parameters.

[0011] Further, in the model encryption and privacy protection method for artificial intelligence algorithms according to the present invention, the step S107 includes: Perform quantum kernel mapping on the user request features through a quantum support vector machine. When it is detected that the number of requests within a unit time exceeds a preset threshold, it is determined as a high-frequency query abnormal behavior; After triggering the defense mechanism, dynamically adjust the interval boundary value according to the ROC curve of the preset model validation set, and convert the probability value output by the updated global model into interval labels of high risk, medium risk, and low risk; Input the abnormal behavior log into the rule matching module of the compliance routing engine to block the transmission of illegal data streams including sensitive fields.

[0012] Furthermore, for the model encryption and privacy protection method for artificial intelligence algorithms according to the present invention, the step S108 includes: Perform digital signature on the federated learning node selection record and noise injection operation log based on the lattice cryptography-based quantum-secure hash algorithm; Synchronize the signed log data to the permissioned blockchain node to generate an immutable audit chain including timestamps; Input the privacy leakage risk metrics in the audit chain into the optimization objective function of the quantum hybrid optimization framework to drive the dynamic adjustment of the encryption strength parameters.

[0013] Furthermore, for the model encryption and privacy protection method for artificial intelligence algorithms according to the present invention, the step S109 includes: Search for the optimal combination of the key length of quantum fully homomorphic encryption and the differential privacy noise variance through the quantum approximate optimization algorithm; Combined with the classical gradient descent algorithm, locally optimize the weight allocation strategy for federated node selection to generate updated encryption strength parameters and noise injection rules; Synchronize the encryption strength parameters and noise injection rules to the aggregation controller and differential privacy module of the federated learning framework in real time.

[0014] Furthermore, for the model encryption and privacy protection method for artificial intelligence algorithms according to the present invention, it further includes: According to the privacy leakage risk score in the audit chain, reversely adjust the selection priority of the federated learning participating nodes; Based on the real-time monitored model accuracy loss data, optimize the differential privacy noise variance and the geographical transmission constraints in the compliance routing rules; Feed back the adjusted node selection priority, noise variance, and geographical transmission constraint parameters to the dimensionality reduction processing unit of the quantum feature extraction module and the key manager of the hierarchical encryption module.

[0015] Advantages of the present invention; Through the quantum optimization node selection and multi-objective weighted aggregation strategy, the present invention effectively solves the parameter conflict problem caused by data heterogeneity in federated learning. The quantum annealing algorithm screens data complementary nodes to reduce parameter divergence, and the weight coefficients optimized by the NSGA-II algorithm balance the contribution differences of heterogeneous data, improving the model convergence efficiency. The hierarchical encryption and adaptive noise regulation mechanism, based on quantum fully homomorphic encryption to protect highly sensitive parameters, combines a dynamic noise injection strategy driven by reinforcement learning to achieve a dynamic balance between privacy protection intensity and model accuracy, reducing the negative impact of noise on convergence stability. The audit-driven closed-loop feedback system reversely optimizes node selection and noise rules through a quantum security audit chain, and the quantum hybrid optimization framework synchronously tunes the encryption intensity and noise parameters, forming a full-link adaptive collaboration from data preprocessing to model inference, ultimately achieving the technical effect of global optimality of model aggregation efficiency and privacy-utility in non-i.i.d. data scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the drawings.

[0017] Figure 1 It is a flowchart of a model encryption and privacy protection method for artificial intelligence algorithms provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions of the present invention in conjunction with the specific embodiments and corresponding drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. The following will detail the technical solutions provided by each embodiment of the present invention in conjunction with the drawings. To better understand the objectives of the present invention, the following will give a more detailed description of the present invention.

[0019] Please refer to Figure 1 , the model encryption and privacy protection method for artificial intelligence algorithms provided by the present invention includes: Step S101, obtaining the original business data, identifying sensitive attributes through quantum feature extraction and dimensionality reduction processing, and generating a dimensionality-reduced feature vector; After obtaining the original business data, the data is preprocessed by standardization to eliminate the influence of dimension differences and outliers. Specifically, it includes performing word segmentation, entity recognition, and missing value imputation on unstructured data, and performing normalization and feature encoding conversion on structured data. The preprocessed data is loaded into the quantum computing unit through a quantum random access memory, and the classical feature vector is mapped into a superposition form of quantum states by using quantum amplitude encoding to form the input data of the quantum principal component analysis algorithm.

[0020] The quantum principal component analysis module calculates the eigenvalues and eigenvectors of the feature covariance matrix in parallel through the quantum phase estimation algorithm. In the quantum circuit, a quantum simulator of the covariance matrix is constructed through controlled rotation gate operations, and the matrix diagonalization process is accelerated by using quantum parallelism. The eigenvectors are arranged in descending order of eigenvalues, and the feature dimensions with a correlation exceeding a preset threshold with the sensitive attributes are screened to generate a low-dimensional feature subspace. The sensitive attribute recognition is based on the calculation of the Pearson correlation coefficient between the eigenvectors and the preset sensitive fields, and the white list of sensitive attributes is defined in combination with the business scenario to dynamically adjust the feature screening conditions.

[0021] The reduced-dimensional eigenvectors are converted into the classical data format through quantum state measurement, retaining the key sensitive attribute correlation dimensions and removing redundant features. The dimension of the eigenvectors is dynamically determined according to the variance interpretation rate of the quantum principal component analysis, and the information retention amount is maximized within the preset compression rate range. The output eigenvectors carry the key sensitive attribute distribution patterns of the original data, providing a structured input for subsequent privacy grading. This process accelerates the processing of high-dimensional data through quantum computing, identifies sensitive information at the feature space construction stage, and reduces the computational complexity of the subsequent privacy protection module.

[0022] After the reduced-dimensional eigenvectors are generated, format verification and data integrity verification are performed to prevent data distortion caused by quantum noise during the quantum computing process. The eigenvectors passing the verification are stored in the distributed database, and a data channel is established with the privacy level mapping table construction module to support the dynamic grading of sensitive levels by the multi-objective optimization algorithm. The collaborative design of preprocessing and feature extraction enables the identification of sensitive attributes and dimension compression of the original business data before entering the federated learning framework, laying a data foundation for end-to-end privacy protection.

[0023] Step S102: Input the reduced-dimensional eigenvectors into a multi-objective optimization and grading algorithm to generate a privacy level mapping table including the data sensitive levels and the cleaned structured data; After the dimensionality-reduced feature vectors are input into the multi-objective optimization and grading module, a set of multi-objective functions including data sensitivity scores, business availability weights, and compliance constraints is constructed to initiate the grading algorithm process. The data sensitivity scores are calculated based on the association strength between the feature vectors and the preset sensitive attributes. The business availability weights are determined by the contribution of the feature dimensions to the prediction of business metrics. The compliance constraints are dynamically loaded according to the data protection regulations in the target area. The module uses an improved NSGA-II algorithm to parallelly evaluate candidate solutions on quantum computing units, generating a set of Pareto front solutions covering different optimization objective equilibriums.

[0024] After generating the Pareto optimal solution set, a multi-criteria decision-making analysis process is executed. Through the preset data minimization principle and business priority rules, the optimal solutions that meet the privacy-utility balance point are screened from the solution set. The selected solutions drive the construction process of the privacy level mapping table, dividing the data fields into three sensitive levels: public level, internal level, and confidential level. The mapping table uses structured tags to label the encryption levels and quantum key parameters of the fields. Among them, the confidential-level fields are associated with high-dimensional sensitive attributes and bound with enhanced encryption policies. This table is distributed to each node through the federated learning control plane to guide subsequent differential encryption processing.

[0025] The generation process of the cleaned structured data is executed synchronously with the privacy grading. According to the sensitive level markings in the privacy level mapping table, field-level filtering and desensitization operations are performed on the original business data. The public-level fields retain the original data format, the internal-level fields perform hash masking processing, and the confidential-level fields apply quantum homomorphic encryption algorithms to transform the data form. The cleaned data retains the business availability characteristics while eliminating sensitive information that directly identifies personal identities, forming a structured data set that complies with the data compliance standards in the target area.

[0026] After receiving the cleaned data stream, the structured data storage module performs distributed sharding storage and access permission configuration. The public-level data is opened to the public storage area of the federated learning nodes. The internal-level data is bound with node identity certificates to achieve fine-grained access control. The confidential-level data is transmitted to trusted nodes through the quantum key distribution optical channel. The data storage format and the privacy level mapping table establish a dynamic association mechanism to support the on-demand invocation of data subsets at different sensitive levels during subsequent training processes. The hierarchical storage strategy triggers the adaptive adjustment of encryption intensity during the data invocation stage, forming a data processing link linked to the privacy level.

[0027] The privacy level mapping table and the structured data are synchronously transmitted to the aggregation controller through the metadata interface of the federated learning framework. The field-level tags of the mapping table provide a benchmark for subsequent adjustment of the encryption strength of model parameters, and the cleaned structured data provides compliant inputs for local training of participating nodes. The collaborative design of the multi-objective optimization grading module and the data cleaning module enables the dynamic adaptation of feature space partitioning and privacy protection strategies during the data preprocessing stage, laying a hierarchical control foundation for the privacy-utility balance of the entire federated learning process.

[0028] Step S103: Based on the privacy level mapping table, dynamically adjust the aggregation weights of the participating nodes in the federated learning framework, and select a subset of low-latency nodes through quantum optimization to generate global model parameters. After the privacy level mapping table is connected to the federated learning framework, the aggregation controller starts a dynamic weight adjustment mechanism. Based on the Gaussian process model constructed by the Bayesian optimizer, collect the historical aggregation contribution degree indicators of each participating node, including the parameter update amplitude, convergence stability, and data distribution representativeness parameters. The model fits the probability distribution characteristics of the node contribution degree through the kernel function and predicts the weight allocation ratio for the next training cycle. The weight prediction results are updated to the federated aggregation control center in real time to form the decision basis for the dynamic weighting strategy.

[0029] The node selection optimization module synchronously receives the communication delay measurement data and the data distribution similarity matrix of the participating nodes. The communication delay parameter is converted into the coupling strength coefficient in the Ising model after normalization, and the data distribution similarity is mapped into the spin interaction energy by calculating the KL divergence. The constructed Hamiltonian is input into the quantum annealing machine, and the quantum tunneling effect is used to break through the local optimal solution limit to solve the index list of the federated node subset that satisfies the delay constraint and has complementary data distributions. This process accelerates the search efficiency of the node combination space through quantum parallel computing and screens out a node cluster with low communication overhead and high data heterogeneity.

[0030] The federated aggregation controller synthesizes the dynamic weight prediction value and the quantum optimization node subset index to reconstruct the weighting strategy of the federated average algorithm. The weight coefficients are dynamically normalized according to the real-time contribution degree of the members within the node subset, and at the same time, a data volume ratio factor is introduced to balance the sample distribution deviation. The adjusted aggregation formula performs a weighted summation operation in the ciphertext state, and combines the sensitive levels marked by the privacy level mapping table to perform hierarchical aggregation on the local model parameters with different encryption strengths. The aggregation result is broadcast to all participating nodes through the quantum secure channel to form a global parameter update amount that takes into account both data privacy and model convergence efficiency.

[0031] After the global model parameters are generated, gradient verification and parameter legality verification are performed. The verification mechanism compares the integrity of the contribution parameters of each node based on the quantum-secure hash algorithm, identifies abnormal update amounts, and triggers the re-aggregation process. The parameter set that passes the verification is input into the model version management module to generate a global model snapshot with a timestamp and a node subset identifier. The snapshot data is synchronized to the federated learning network through the distributed storage interface, supporting participating nodes to start the next round of local training by calling the latest model parameters as needed. The co-design of dynamic weight adjustment and quantum optimization node selection effectively alleviates the parameter drift caused by non-independent and identically distributed data and improves the model convergence stability in heterogeneous environments.

[0032] In the global parameter distribution phase, the aggregation controller enables the quantum key distribution optical channel transmission for the high-sensitivity layer parameters according to the encryption level label in the privacy level mapping table. The low-sensitivity layer parameters are transmitted using the classical encryption protocol to reduce communication resource consumption. The coordinated implementation of the hierarchical transmission strategy and the dynamic aggregation mechanism realizes the adaptive balance between the privacy protection intensity and the communication efficiency in the federated learning process, forming a full-link optimization closed-loop from node selection, weight allocation to parameter generation.

[0033] Step S104, perform quantum fully homomorphic encryption on the high-sensitivity layer in the global model parameters according to the privacy level mapping table, and perform quantum sparse compression on the encrypted parameters; After receiving the global model parameters, the privacy level mapping table parsing module identifies the high-sensitivity layer parameters based on the preset sensitive layer labels, including key data such as the output layer weight matrix and the user embedding vector. The identification process uses a field-level label matching algorithm to divide the parameter tensor into data blocks of different sensitive levels, and the high-sensitivity layer data blocks trigger the encryption processing instruction. The encryption instruction is distributed to the quantum fully homomorphic encryption engine through the federated learning control plane to start the parameter protection process.

[0034] The quantum fully homomorphic encryption engine performs encryption operations based on lattice cryptography on the high-sensitivity layer parameters. The encryption process uses the public key generated by the quantum key distribution protocol to perform homomorphic conversion on the parameters, supporting addition and multiplication operations in the ciphertext state. After the parameter tensor completes the encryption operation in the quantum state, it retains its mathematical operation characteristics, forming an encrypted parameter set that can directly participate in the calculation during the federated learning aggregation process. The encrypted parameters are transmitted to the sparse compression module through the quantum communication channel to avoid the risk of privacy leakage caused by plaintext transmission.

[0035] After receiving the encryption parameters, the quantum sparse compression module uses the encoder unit of the quantum variational autoencoder to perform low-dimensional manifold projection on the ciphertext data. The encoder quantum circuit maps the high-dimensional parameter matrix to a low-dimensional sparse space by optimizing the sparse basis vector combination, and the volume of the compressed ciphertext is significantly reduced. During the compression process, the continuity of the encryption state is maintained, so that the ciphertext data is compatible with the subsequent homomorphic operation interface before and after compression. The sparse coding parameters are temporarily stored in the quantum random access memory and wait to be input into the differential privacy noise injection module.

[0036] Before the compressed ciphertext parameters are output, format compatibility check and data integrity verification are performed. The verification mechanism compares the data fingerprints before and after compression based on the quantum secure hash algorithm to identify data distortions caused by quantum noise or transmission errors. The verified compressed ciphertext parameters are transmitted to the noise injection module through the homomorphic operation interface, and at the same time, the compression log is recorded in the audit trace database. The collaborative design of the compression module and the encryption module reduces the communication overhead while maintaining data privacy, forming a closed-loop protection mechanism for the high-sensitive parameter processing link.

[0037] The compressed encryption parameters are stored in the distributed ciphertext database and dynamically associated with the privacy level mapping table. The storage module configures access permissions according to the sensitive level label. Compressed ciphertext at the high-sensitive level is only allowed to be called by authorized nodes through the quantum secure channel. The storage strategy is connected with the subsequent model update process, supporting fast positioning of encrypted compressed data blocks during incremental parameter replacement, and improving the global model iteration efficiency. The dual processing mechanism of encryption and compression provides a full-process privacy protection foundation for the transmission, storage, and calculation of high-sensitive data in the federated learning framework.

[0038] Step S105, injecting dynamic noise generated based on reinforcement learning into the compressed encryption parameters to generate encrypted model parameters with noise; After the compressed encryption parameters are input into the noise injection module, the deep deterministic policy gradient reinforcement learning model starts the dynamic noise policy generation process. The model receives the real-time monitoring data stream of the federated learning system, including the remaining privacy budget, the accuracy loss of the model validation set, and the magnitude of the parameter update gradient. The state space is constructed as the time series features of the privacy-utility balance index, and the action space is defined as the set of layer position indexes and variance intensity parameters of noise injection. The reward function combines the negative constraint of the privacy budget consumption rate and the weight of the model accuracy loss, driving the policy network to generate a dynamic noise configuration scheme.

[0039] The quantum random number generator is based on the physical entropy source of quantum randomness of single photon detectors, and the original random sequence is fitted with a Gaussian distribution through a post - processing algorithm. After the generated noise sequence is verified for its distribution characteristics by the Kolmogorov - Smirnov test, it is input into the buffer queue of the noise injection controller. The dimension of the noise sequence is automatically aligned with the tensor structure of the encryption parameters, realizing a one - to - one mapping relationship between the noise elements and the target parameters, and avoiding the perturbation of the model accuracy caused by the injection position deviation.

[0040] After receiving the encryption parameters and the noise sequence, the homomorphic noise superposition interface performs the homomorphic addition gate operation defined by the quantum fully homomorphic encryption protocol. The encryption parameters are linearly superposed with the noise sequence in the quantum state, and the integrity of the ciphertext format is verified through the quantum Fourier transform during the superposition process. The injected noisy parameters retain the homomorphic operation ability and satisfy the differential privacy constraint conditions after decryption, forming an encrypted noise parameter set that takes into account both privacy protection and model usability. The noise injection log is recorded in the audit tracking database in real - time, providing a data basis for subsequent policy optimization.

[0041] After the noisy encryption parameters are generated, noise impact assessment and model convergence monitoring are performed. The evaluation module compares the change amplitude of the model gradient before and after the noise injection, identifies abnormal parameter fluctuations and triggers dynamic calibration of the noise variance. The calibration signal is fed back to the reinforcement learning policy network to optimize the noise configuration parameters in the next training cycle. The monitoring data is synchronously input into the federated learning aggregation controller, providing a noise compensation reference value for the global model parameter update and reducing the negative impact of noise injection on the model convergence rate.

[0042] The noise parameters are transmitted to the distributed storage nodes through a quantum - secure channel and associated indexes are established with the privacy level mapping table. The storage module configures multi - version noise parameter snapshots according to the sensitive level labels, supporting the on - demand loading of historical noise configurations when the model is rolled back. The collaborative design of the dynamic noise injection mechanism and the encryption compression module realizes the adaptive balance between the privacy protection intensity and the model utility while maintaining the continuity of ciphertext operations, providing data integrity guarantee for subsequent incremental updates and audit tracing.

[0043] Step S106, locate the incrementally updated part of the noisy encrypted model parameters through quantum hash comparison, generate an increment list and update the global model; After the noisy encrypted model parameters are input into the quantum hash comparison module, quantum state encoding and data difference location of the historical version parameters are performed. The quantum circuit constructed by the Grover algorithm is used to compare the current parameters with the historical version in parallel, and the quantum amplitude amplification principle is used to accelerate the search for parameter differences. When the detected parameter value deviation exceeds the preset threshold, it is marked as an incrementally updated item, and an increment matrix containing the parameter index and the difference value is generated. This process significantly improves the comparison efficiency of large - scale parameter sets through quantum parallel computing.

[0044] After the incremental matrix is input into the sparse coding module, the quantum variational autoencoder performs low-rank approximation decomposition on the differential data. The encoder quantum circuit learns the sparse representation pattern of the incremental parameters and projects the high-dimensional differential matrix into the low-dimensional space spanned by the orthogonal basis vectors. The compressed sparse representation data is re-encrypted through a quantum fully homomorphic encryption scheme based on learning with errors on rings to generate compressed ciphertext data packets that meet the data security standards. The compression process maintains the ciphertext operation characteristics and supports the homomorphic processing requirements in the subsequent aggregation stage.

[0045] After the compressed ciphertext data packets are transmitted to the blockchain storage interface of the audit trail module, a smart contract is called to generate an audit transaction request containing the timestamp and the operator identifier. The transaction request triggers the permissioned blockchain consensus node to execute the zero-knowledge proof verification process to verify the legality of the incremental parameter modification and the compliance of the operation permissions. The verified incremental parameters are decrypted and merged into the global model copy to generate an updated model parameter set with a version identifier. The update log generates a digital fingerprint through a quantum-secure hash algorithm and is synchronously written into the blockchain to form an immutable audit chain record.

[0046] After the global model update is completed, the version management module pre-sets the parameter rollback ability. A rollback index table is established through the multi-version parameter snapshots stored in the blockchain. When it is detected that the model accuracy drops abnormally, the smart contract is triggered to load the historical parameter versions in time series. The rollback mechanism and the incremental update process form a closed loop, providing fault tolerance guarantee for model iteration. The updated parameter set is synchronized to the federated learning participating nodes through the quantum key distribution optical channel, supporting each node to start the local training task based on the latest model.

[0047] The collaborative design of the incremental update process and the quantum hash comparison module reduces the communication overhead while maintaining data integrity. The connection between the compression and encryption mechanism and the blockchain verification link realizes the full-process traceability from difference location to model update. The dynamic parameter replacement strategy combined with the version rollback ability improves the robustness and iteration efficiency of the federated learning framework in the non-independent and identically distributed data scenario, forming a continuous optimization link for privacy protection and model utility.

[0048] Step S107, receive the user inference request, use quantum anomaly detection to identify high-frequency query behaviors, and perform fuzzification processing on the output results of the updated global model; After the user's reasoning request is connected to the federated learning framework, the quantum anomaly detection module starts the real-time request feature analysis process. The request feature vector is mapped to the high-dimensional Hilbert space by the quantum kernel function, and the query behavior pattern is determined by the classification boundary constructed by the quantum support vector machine. When the request frequency in the unit time window exceeds the dynamic threshold, the abnormal behavior marking mechanism is triggered. The dynamic threshold is dynamically adjusted according to the historical access pattern baseline and the compliance constraints preset by the target area data protection regulations to meet the detection sensitivity requirements of different business scenarios.

[0049] After detecting high-frequency query anomalies, the model output adaptation module performs fuzzification processing on the global model inference results. The probability value output is converted into discrete risk interval labels through quantum fuzzy logic operations, including high risk, medium risk and low risk levels. The interval boundary values ​​are dynamically calibrated based on the ROC curve of the model verification set to balance the detection sensitivity and false alarm rate indicators. The converted interval labels are returned to the user terminal through the quantum secure channel, hiding the original probability distribution details to reduce the possibility of privacy leakage.

[0050] Abnormal behavior logs are input into the rule matching module of the compliance routing engine in real time, and the preset list of sensitive fields and cross-border data transmission policies are loaded. When abnormal requests are identified to contain user ID numbers, biometrics and other tag fields, the traffic filtering rule set is activated. The interception instruction triggers the quantum secure hash engine to generate an operation fingerprint, which is bound to the abnormal event log and written into the blockchain audit chain to provide reverse data flow for subsequent model optimization and policy adjustment. The blocked data flow is redirected to the isolated storage area, waiting for the manual review process to process.

[0051] The compliance routing engine matches the transmission rule base according to the data regional label, and generates interception instructions for requests that violate the data localization requirements of the target area. The transmission rule base integrates geo-fence parameters and real-time policy update data, and dynamically loads applicable terms through the semantic parsing engine. The interception operation record generates a verifiable credential through the quantum security signature algorithm, which is synchronized to the permission chain node to form an unalterable audit evidence chain. The collaborative design of the routing strategy and the anomaly detection module superimposes transmission control on the basis of output fuzzification to form a multi-level privacy protection system.

[0052] The output data after fuzzification is temporarily stored through the distributed cache interface to perform access frequency monitoring and data lifecycle management. The cache module configures differentiated retention periods according to the sensitivity level labels, and enables a short-term automatic erasure mechanism for high-risk label data. A dynamic association is established between the data flow path and the privacy level mapping table to maintain business continuity while reducing storage risks. The closed-loop design of anomaly detection and output processing enables the user-side reasoning service to build a full-link privacy protection network from request access to result feedback while meeting low-latency response requirements.

[0053] Step S108, intercept the illegal transmission according to the data regional label and match the compliance rules, and generate an audit chain through quantum-secure hashing; After receiving the transmission request, the data regional label matching module calls the geofencing policy library to load the data compliance rules for the target area. The geofencing parameters are dynamically generated based on the data storage location, transmission destination, and governing regulations. The natural language clauses are converted into executable policy instructions through the semantic parsing engine. The policy library integrates the regionally updated data protection regulations in real time, including data localization requirements and cross-border transmission whitelists, to form a multi-dimensional set of compliance constraints.

[0054] The transmission request parsing engine extracts the regional label and content features in the packet metadata and performs multi-level matching with the compliance rule library. The matching process uses a rule determination algorithm based on a finite state machine. When it detects that the data flow is to a non-whitelist area or contains restricted sensitive fields, a hierarchical interception mechanism is triggered. The primary interception performs packet filtering at the protocol layer, and the advanced interception activates the session termination instruction at the application layer to form a multi-level transmission control system. After the interception instruction is generated, the audit log recording process is started synchronously.

[0055] The quantum-secure hashing engine performs quantum-resistant digital signatures on the intercepted operation logs. The hashing function based on lattice cryptography converts the operation type, timestamp, and data fingerprint into a fixed-length digest, and uses the private key generated by the quantum key distribution protocol for signing. The signed log data is submitted through the permission chain node consensus interface, and after being verified by the practical Byzantine fault tolerance mechanism, it is appended to the blockchain data structure to generate an immutable audit chain with a time series. The audit chain block header information contains a forward hash pointer and a quantum-secure signature, forming a cross-block data integrity verification link.

[0056] The audit chain storage module distributes and stores the signed logs in the encrypted database of the permission chain nodes, and establishes access control based on the sensitive level. The high-sensitivity operation records enable the quantum key distribution optical channel transmission protocol, and only authorized audit nodes can access the complete log data through the quantum-secure proxy. The storage architecture and the privacy level mapping table establish a dynamic association index to support multi-dimensional rapid retrieval of audit information according to regional labels, operation types, etc.

[0057] The audit chain analysis module periodically scans the blockchain ledger to extract risk indicators such as the frequency of illegal transmission events and the exposure duration of sensitive data. The indicator data is input into the objective function of the quantum hybrid optimization framework to drive the dynamic optimization of subsequent encryption intensity and node selection strategies. The reverse feedback mechanism of the illegal log is connected to the real-time update interface of the policy library to form a closed-loop optimization link from transmission control to rule iteration. The collaborative design of regional compliance policies and quantum audit chains constructs a verifiable data governance system throughout the life cycle while blocking illegal transmissions.

[0058] Step S109: Based on the risk indicators in the audit chain, dynamically adjust the encryption strength and noise injection strategy, and feedback the adjusted policy parameters to the federated learning node selection and encryption module to form a closed-loop iteration.

[0059] The audit chain risk indicator extraction module periodically scans the blockchain ledger and analyzes quantitative parameters such as the frequency of illegal transmission events, the duration of sensitive data exposure, and the node trustworthiness score. After normalization, the risk indicators are input into the quantum hybrid optimization framework to construct a multi-objective optimization function centered on the balance between privacy protection strength and model utility. The quantum approximate optimization algorithm parallelly searches for the optimal combination of the key length of quantum fully homomorphic encryption and the differential privacy noise variance on a quantum processor to generate a candidate parameter set that meets the current risk level.

[0060] The classical optimization module extracts the feasible solution interval from the quantum optimization results, and combines the real-time communication efficiency and data distribution characteristics of the federated learning nodes to iteratively adjust the node selection weight allocation strategy through the gradient descent algorithm. The weight update signal is fed back to the quantum optimization module through the quantum-classical hybrid interface to dynamically shrink the boundary of the parameter search space and improve the convergence speed of the hybrid optimization process. The optimized encryption strength parameters and noise rules form a policy update package, which is pushed to the federated learning control plane after digital signature verification.

[0061] After receiving the update package, the policy distribution engine starts the parameter hot loading process. The encryption strength parameters are written into the key scheduler of the quantum fully homomorphic encryption module, triggering key rotation and encryption protocol version update. The noise injection rules are updated to the policy cache area of the differential privacy module to reconfigure the noise variance threshold and injection position priority. The node selection weight parameters are synchronized to the federated aggregation controller to reconstruct the weighted aggregation formula for the next training cycle. The parameter update log generates an audit fingerprint through the quantum secure hash algorithm and is appended to the blockchain to form a policy iteration traceability link.

[0062] The closed-loop feedback mechanism reversely inputs the optimized policy parameters into the data preprocessing stage. The dimensionality reduction unit of the quantum feature extraction module receives the updated regional transmission constraint parameters and dynamically adjusts the weight coefficient of the geographical label in feature screening. The key manager of the hierarchical encryption module updates the quantum key distribution policy according to the node priority table and enables an enhanced anti-quantum attack protocol for high-priority communication links. The coordinated optimization of the full-link parameters enables the data preprocessing, model training, and inference links to form a dynamically linked privacy-utility balance system.

[0063] The policy iteration verification module monitors the actual effects of the encryption intensity and noise policy in the new cycle of federated learning. By comparing the newly generated privacy risk metrics in the audit chain with the historical baseline data, it evaluates the impact of parameter adjustment on the model convergence speed and the probability of privacy leakage. The evaluation results are input into the quantum hybrid optimization framework as feedback signals to drive the dynamic tuning of the next round of policy parameters. The closed-loop iteration mechanism maintains the privacy protection intensity and model availability of the federated learning system in the data heterogeneous scenario through continuous risk perception and policy response.

[0064] In the data processing stage, the original business data is input into the quantum feature extraction module. Through the quantum principal component analysis algorithm, the feature covariance matrix is calculated in parallel to identify sensitive attributes including user identity information and location coordinates, and the dimensionality-reduced feature vector is output. This process uses the parallel computing characteristics of qubits to accelerate the diagonalization of the covariance matrix, and the dimensionality of the extracted feature vector is reduced by more than 60% compared with the classical algorithm, providing an input basis for subsequent privacy grading.

[0065] The dimensionality-reduced feature vector is input into the multi-objective optimization grading module. With the data sensitivity score, business availability metrics, and GDPR compliance requirements as optimization objectives, the Pareto front solution set is generated through the NSGA-II algorithm. According to the solution set, a privacy level mapping table is constructed to divide the data fields into public layer, internal layer, and confidential layer, and at the same time, the cleaned structured data is output. The privacy level mapping table labels the sensitive levels through field-level tags to guide the differential encryption policies of the federated learning nodes.

[0066] After receiving the privacy level mapping table, the federated learning framework constructs a Gaussian process model based on the Bayesian optimizer during the local training process of the participating nodes to predict the aggregation weights according to the historical contribution degree of the nodes. At the same time, the node communication delay and data distribution similarity parameters are mapped to the Hamiltonian of the Ising model, and the optimal node subset is solved through the quantum annealing machine. The dynamically adjusted aggregation weights and the filtered low-latency nodes are jointly input into the aggregation formula to generate the global model parameters.

[0067] The global model parameters enter the hierarchical encryption module. According to the confidential levels marked in the privacy level mapping table, the quantum fully homomorphic encryption algorithm is applied to the output layer parameters. The encrypted parameters are input into the quantum neural network compression module, and the ciphertext volume is compressed by more than 70% through sparse coding. The compressed ciphertext parameters are transmitted to the noise injection interface while maintaining the encrypted state to avoid the data leakage risk in the intermediate decryption link.

[0068] After the noise injection module receives the encryption parameters, the deep deterministic policy gradient model generates the model layer position and variance intensity policy for noise injection based on the real-time privacy budget consumption rate and the model accuracy loss data. The quantum random number generator generates a true random noise sequence that conforms to the Gaussian distribution, and superimposes the noise onto the ciphertext parameters through quantum homomorphic addition operations. The parameters after noise injection still satisfy the differential privacy constraint after decryption processing, forming encrypted model parameters with noise.

[0069] Based on the Grover algorithm, the incremental update module compares the parameters with noise with the historical version in the quantum state to locate the newly added or modified parameters and generate an incremental list. After being compressed by quantum sparse coding, the incremental list is re-encrypted through the quantum fully homomorphic encryption algorithm and input into the blockchain storage node. The global model replaces the parameters according to the incremental list to generate an updated model version, which is synchronized to each federated learning node.

[0070] When a user initiates an inference request, the quantum support vector machine performs a quantum kernel space mapping on the request features to detect high-risk queries whose request frequency exceeds a preset threshold within a unit time. After the defense mechanism is triggered, the precise probability value output by the model is converted into high, medium, and low risk interval labels through quantum fuzzy logic. The abnormal query log is input into the compliance routing engine in real time to block the cross-regional transmission of illegal data including sensitive fields.

[0071] The compliance routing engine matches the preset transmission rules according to the data geographical region label, and generates an interception instruction for requests that violate the data localization requirements in Region A. All operation logs are digitally signed through a quantum-secure hash algorithm based on lattice cryptography to generate an audit chain data block including a timestamp. The audit chain data is stored on the chain in real time to form an immutable operation trace record.

[0072] The audit chain data is input into the quantum hybrid optimization framework. The quantum approximate optimization algorithm takes the privacy leakage risk score and the model accuracy index as the optimization objectives to search for the optimal combination of the encryption key length and the noise variance. The classical gradient descent algorithm synchronously optimizes the weight allocation parameters in the federated node selection strategy. The updated encryption intensity parameters and noise rules are synchronously updated to the dimensionality reduction unit of the feature extraction module and the key manager of the encryption module in real time, forming a technical closed loop from data input to policy iteration.

[0073] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, the step S101 includes: Parallelly calculate the feature covariance matrix through quantum principal component analysis, and extract the feature vectors associated with sensitive attributes; Input the feature vectors into the multi-objective optimization hierarchical algorithm, and simultaneously take the data sensitivity classification, business availability score, and GDPR compliance index as the optimization objectives to generate a Pareto optimal solution set; Generate a privacy level mapping table based on the optimal solutions in the Pareto optimal solution set. The privacy level mapping table includes encryption level tags for each data field, which are used to control the encryption intensity of the local training parameters of the federated learning participating nodes.

[0074] In the quantum feature extraction stage, after receiving the original business data, the quantum principal component analysis module diagonalizes the feature covariance matrix using the quantum bit parallel computing mechanism. This process accelerates the eigenvalue decomposition through the quantum phase estimation algorithm, identifies the sensitive attribute dimensions related to user identity information and location trajectories, and generates a set of low-dimensional feature vectors. The parallel nature of quantum computing improves the high-dimensional data processing efficiency by two orders of magnitude compared with classical algorithms, providing a high-precision feature space representation for subsequent privacy grading.

[0075] After receiving the dimensionality-reduced feature vectors, the multi-objective optimization grading module constructs a multi-objective function including data sensitivity scores, business availability weights, and compliance constraints in Region A. The data sensitivity scores are calculated based on the Pearson correlation coefficient between the feature vectors and sensitive attributes, and the business availability weights are determined by the regression contribution of the feature dimensions to the business metrics. The improved NSGA-II algorithm is used to parallelly evaluate the solution set on the quantum computing unit to generate a set of Pareto front solutions covering different optimization objective balance points. Each Pareto solution corresponds to a specific privacy-utility balance strategy.

[0076] The privacy level mapping table construction module performs multi-criteria decision analysis on the Pareto optimal solution set and selects the optimal solution plan that meets the data minimization principle in Region A. According to the selected plan, the data fields are divided into three sensitive levels: public level, internal level, and confidential level. The confidential level fields include the feature dimensions that directly identify personal identities. The mapping table is stored in the XML structured format, and each field is marked with an encryption level label and the corresponding quantum key length parameter. This table is distributed to each participating node through the federated learning control plane to guide the differential encryption processing of parameters at different sensitive levels during the local training process.

[0077] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, step S103 includes: Construct a Gaussian process model based on the Bayesian optimizer, and predict the weight values of each federated learning participating node according to the historical aggregation contribution degrees of each federated learning participating node; Map the node communication delay and data distribution similarity to the Hamiltonian of the Ising model, and solve the optimal federated node subset through a quantum annealer; According to the predicted weight values of each federated learning participating node and the optimal federated node subset, adjust the weighted average coefficient in the federated aggregation formula to generate global model parameters.

[0078] In the stage of predicting the weights of the federated learning nodes, the Bayesian optimizer collects the model update contribution data of each participating node in the previous training rounds, including the parameter update amplitude and the convergence stability index. The Gaussian process model uses the historical contributions as training samples, fits the probability distribution characteristics of the node contributions through the kernel function, and outputs the expected weight values for future training rounds. The weight prediction results are updated in real time to the federated aggregation control center, providing a prior knowledge base for dynamic adjustment.

[0079] The node selection optimization module synchronously receives the communication delay measurement data of the participating nodes and the data distribution similarity matrix. The communication delay data is normalized and then transformed into the coupling strength coefficient in the Ising model, and the data distribution similarity is mapped to the spin interaction energy through the KL divergence calculation. The constructed Hamiltonian of the Ising model is input into the quantum annealing machine, which quickly converges to the ground state with the assistance of the quantum tunneling effect, and outputs the optimal node subset index list that satisfies the delay constraint and has complementary data distributions.

[0080] The federated aggregation controller improves the traditional federated averaging algorithm by integrating the weight prediction values and the node subset indices. The weighted average coefficient is dynamically normalized according to the real-time weight values of the members within the node subset, and at the same time, a data volume proportion factor is introduced to balance the sample distribution deviation. The adjusted aggregation formula performs the weighted summation operation in the ciphertext state to generate the global model parameters that meet the data compliance requirements of Region A. These parameters are broadcast to all participating nodes through the quantum secure channel to start the next round of the federated training cycle.

[0081] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, the step S104 includes: Encrypt the global model parameters marked as the high-sensitivity level in the privacy level mapping table using the quantum fully homomorphic encryption algorithm; Compress the encrypted parameters through the sparse coding layer of the quantum neural network to generate compressed ciphertext parameters; Input the compressed ciphertext parameters into the homomorphic operation interface of the differential privacy noise injection module to maintain the noise superposition in the ciphertext state.

[0082] In the model parameter encryption stage, the privacy level mapping table parsing module identifies the model parameter levels marked as confidential, including the user portrait embedding vector and the output layer weight matrix. The quantum fully homomorphic encryption engine adopts an encryption scheme based on lattice cryptography, assigns independent quantum key pairs to each confidential parameter, and performs addition and multiplication homomorphic operations in the ciphertext space. The encrypted parameters maintain the tensor data structure and are transmitted to the compression processing module through the quantum channel.

[0083] After receiving the encryption parameters, the compression processing module enables the sparse coding layer in the quantum neural network to learn the low-dimensional manifold representation of the ciphertext data through the quantum variational autoencoder structure. This layer constructs an encoder-decoder architecture using quantum circuits, projecting the high-dimensional parameter matrix onto the sparse basis vector space while keeping the ciphertext state unchanged. The volume of the compressed ciphertext parameters is reduced to less than 30% of the original data, while maintaining the compatibility of homomorphic operations.

[0084] The compressed ciphertext parameters are input into the differential privacy noise injection interface, which conducts a protocol handshake with the quantum homomorphic operation unit to verify the compatibility of the ciphertext format. The noise generator generates a true random noise sequence based on the principle of quantum random walk and superimposes the noise onto the compressed ciphertext parameters through homomorphic addition operations. The entire superimposition process is completed in the ciphertext space, avoiding the risk of information leakage caused by parameter decryption and meeting the requirements of the data security regulations in Region A for the processing of sensitive information.

[0085] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, step S105 includes: Using a deep deterministic policy gradient reinforcement learning model, with the privacy budget consumption rate and model accuracy loss as the reward functions, to generate the model layer position and variance intensity strategy for noise injection; Using a quantum random number generator to generate a noise sequence that conforms to the Gaussian distribution; Through the homomorphic addition operation of quantum fully homomorphic encryption, superimpose the noise sequence onto the ciphertext of the encryption parameters to generate encrypted model parameters with noise.

[0086] In the noise strategy generation stage, the deep deterministic policy gradient reinforcement learning model receives real-time monitoring data from the federated learning system, including the remaining privacy budget, the accuracy of the model validation set, and the magnitude of the parameter update gradient. The state space of the model is defined as the time series characteristics of the privacy-accuracy balance metric, and the action space is the position index and variance parameters of the noise injection layer. The reward function is designed as a weighted combination of the negative logarithmic function of the privacy budget consumption rate and the model accuracy loss, driving the policy network to generate a dynamic noise configuration scheme.

[0087] The quantum random number generation module is based on the physical entropy source of quantum randomness of single-photon detectors and fits the original random sequence to the Gaussian distribution through a post-processing algorithm. After the generated noise sequence is verified for its distribution characteristics by the Kolmogorov-Smirnov test, it is input into the buffer of the noise injection controller. This sequence is automatically aligned with the dimensions of the encryption parameter tensor to ensure a one-to-one correspondence between the noise elements and the parameters to be processed.

[0088] After receiving the encryption parameters and the noise sequence, the homomorphic noise superposition interface performs the homomorphic addition gate operation defined by the quantum fully homomorphic encryption protocol. The encryption parameters are linearly superposed with the noise sequence in the quantum state, and the integrity of the ciphertext format is verified through the quantum Fourier transform. The parameters after noise injection retain the homomorphic operation ability, and after decryption, they satisfy the (ε, δ)-differential privacy constraint conditions and comply with the data anonymization processing standard in Region A. The noise injection log is recorded in real time in the audit tracking database for the subsequent policy optimization module to evaluate the effect.

[0089] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, the step S106 includes: Based on the Grover algorithm, perform quantum state comparison on the current global model parameters and historical versions, locate the newly added or modified parameters, and generate an incremental parameter list; Perform quantum sparse coding processing on the incremental parameter list and generate compressed ciphertext through the quantum fully homomorphic encryption algorithm; Input the compressed ciphertext into the blockchain storage interface of the audit tracking module, trigger the global model update, and generate the updated global model parameters.

[0090] In the quantum hash comparison stage, the Grover algorithm quantum circuit receives the quantum state encoded data of the current global model parameters and historical versions, and accelerates the search for parameter differences through the principle of quantum amplitude amplification. The algorithm compares all parameter components in parallel in the quantum superposition state. When the detected parameter value deviation exceeds the preset threshold, it is marked as a modified item, and an index list of incremental parameters and a numerical difference matrix are generated. This process improves the efficiency by a square root factor compared to the classical comparison algorithm and is suitable for rapid difference localization of large-scale model parameters.

[0091] The incremental parameter list is input into the quantum sparse coding module, and the quantum variational autoencoder is used to perform low-rank approximate decomposition on the difference matrix. The encoder quantum circuit learns the sparse representation pattern of the parameter differences and projects the high-dimensional incremental data into the low-dimensional space spanned by orthogonal basis vectors. The compressed sparse representation data is re-encrypted through a quantum fully homomorphic encryption scheme based on learning with errors over rings to generate a compressed ciphertext data packet that complies with the data security standard in Region A.

[0092] After the blockchain interface of the audit tracking module verifies the compliance of the compressed ciphertext format, it calls the smart contract to generate an audit transaction request including a timestamp. The transaction request triggers the permissioned chain consensus node to execute the model update verification process, and verifies the legality of the incremental parameter modification through zero-knowledge proof. The verified incremental parameters are decrypted and merged into the global model copy to generate an updated model parameter set with a version identifier. The updated parameter set is synchronized to the federated learning participating nodes through the quantum key distribution optical channel to complete the closed-loop management of the model iteration cycle.

[0093] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, step S107 includes: Perform quantum kernel mapping on the user request features through a quantum support vector machine. When it is detected that the number of requests within a unit time exceeds a preset threshold, it is determined as a high-frequency query abnormal behavior; After triggering the defense mechanism, convert the probability value output by the updated global model into interval labels of high risk, medium risk, and low risk; Input the abnormal behavior log into the rule matching module of the compliance routing engine to block the transmission of illegal data streams including sensitive fields.

[0094] In the quantum anomaly detection stage, after receiving the feature vector of the user request, the quantum support vector machine maps the features to a high-dimensional Hilbert space through the quantum kernel trick. The quantum kernel function is constructed using a radial basis function based on the quantum Fourier transform, and the quantum parallelism is used to accelerate the calculation of the kernel matrix. When the detection engine statistically finds that the request frequency of a specific user session within a fixed time window exceeds a dynamic threshold (this threshold is set according to the historical access pattern and the data protection regulations in Region A), an abnormal behavior mark is triggered and a security alarm event is generated.

[0095] After the defense mechanism is triggered, the model output adaptation module performs fuzzification processing on the inference result of the global model. The probability value is mapped to a discrete interval through quantum fuzzy logic operations, where the high-risk interval corresponds to the distribution area of the top 10% of the probability values, the medium-risk interval covers the middle 30% of the numerical range, and the low-risk interval includes the remaining 60% of the data. The interval boundary values are dynamically adjusted according to the ROC curve of the model validation set to maintain the balance between detection sensitivity and false alarm rate.

[0096] After the compliance routing engine receives the abnormal behavior log, the rule matching module loads the preset sensitive field list and data cross-border transmission constraint policies. When it is detected that the abnormal request includes marked fields such as the user's ID number and biometric features, the traffic filtering rule is activated in real time. The interception operation triggers the quantum security hash engine to generate an operation fingerprint, which is bound to the abnormal event log and written into the blockchain audit log to provide a feedback data stream for subsequent model optimization. The blocked data stream is redirected to the isolation storage area through a secure channel and waits for the manual review process to handle.

[0097] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, step S108 includes: Perform digital signature on the federated learning node selection record and noise injection operation log based on the lattice cryptography-based quantum secure hash algorithm; Synchronize the signed log data to the permissioned blockchain node to generate an immutable audit chain including timestamps; Input the privacy leakage risk indicators in the audit chain into the optimization objective function of the quantum hybrid optimization framework to drive the dynamic adjustment of the encryption strength parameters.

[0098] In the audit chain generation stage, the quantum-secure hash algorithm based on lattice cryptography receives the data stream of the federated learning node selection records and the noise injection operation logs. This algorithm constructs a quantum-resistant hash function using the learning with errors (LWE) problem and digitally signs the node identifier, timestamp, and operation type fields in the logs. The signing process is completed on a quantum co-processor to generate a verifiable credential that binds the operation content to the operator's identity, meeting the electronic evidence preservation standard in Region A.

[0099] The signed log data is formatted and packaged into an audit transaction unit and submitted to the blockchain network through the consensus protocol interface of the permissioned chain. The permissioned chain adopts the practical Byzantine fault tolerance consensus mechanism, and the authorized verification nodes perform timestamp sorting and legality verification on the transaction unit. The verified audit transactions are appended to the blockchain data structure in chronological order to form an immutable audit chain with forward hash pointer links and stored in the distributed ledger nodes.

[0100] The privacy leakage risk indicator extraction module in the audit chain periodically scans the blockchain ledger to count the frequency of abnormal node selection events and the deviation degree indicator of noise injection. The quantified value of the risk indicator is input into the multi-objective function of the quantum hybrid optimization framework, which together with the model accuracy loss constitutes the constraint conditions of the Pareto optimization frontier. The optimization result dynamically adjusts the key update period and noise variance threshold of the quantum fully homomorphic encryption, and sends the adjusted parameters to the federated learning aggregator and the noise injection engine through the control plane to complete the adaptive tuning of the privacy protection intensity.

[0101] Specifically, for the model encryption and privacy protection method for artificial intelligence algorithms, step S109 includes: Search for the optimal combination of the key length of the quantum fully homomorphic encryption and the differential privacy noise variance through the quantum approximate optimization algorithm; Combined with the classical gradient descent algorithm, locally optimize the weight allocation strategy for the federated node selection to generate the updated encryption strength parameters and noise injection rules; Synchronize the encryption strength parameters and noise injection rules to the aggregation controller and the differential privacy module of the federated learning framework in real time.

[0102] In the initialization stage of the quantum hybrid optimization framework, the quantum approximate optimization algorithm receives privacy risk metrics from the audit chain and federated learning performance monitoring data. The quantum circuit construction includes a Hamiltonian with key length parameters and noise variance parameters, and searches for the Pareto optimal solution space through the quantum variational optimization loop. The optimization process parallelly evaluates the privacy-utility balance of different parameter combinations on the quantum processor, outputs a candidate parameter set that meets the compliance standards in Region A, and forms an initial optimization strategy library.

[0103] The classical optimization module extracts the feasible solution interval from the quantum optimization results and constructs the constraint condition space for the federated node weight allocation. The gradient descent algorithm takes the node data distribution uniformity and communication efficiency as the optimization objectives, and iteratively adjusts the node selection weight coefficients within the feasible solution interval. The weight update amount of each iteration is fed back to the quantum optimization module through the quantum-classical interface, dynamically shrinking the parameter search range and realizing the collaborative acceleration of the hybrid optimization process.

[0104] After receiving the finally optimized key length, noise variance, and node weight parameters, the parameter synchronization engine writes the encryption strength parameters into the key scheduler of the quantum fully homomorphic encryption module through the policy distribution channel of the federated learning control plane. The noise injection rule update package is pushed to the policy buffer of the differential privacy module after digital signature verification, triggering real-time configuration hot loading. The aggregation controller synchronously receives the node weight parameters and enables the updated weighted aggregation formula in the next round of federated training cycle to complete the closed-loop update of the full-link policy.

[0105] Specifically, the model encryption and privacy protection method for artificial intelligence algorithms further includes: According to the privacy leakage risk score in the audit chain, reversely adjust the selection priority of the federated learning participating nodes; Based on the real-time monitored model accuracy loss data, optimize the geographical transmission constraints in the differential privacy noise variance and compliance routing rules; Feed back the adjusted node selection priority, noise variance, and geographical transmission constraint parameters to the dimensionality reduction processing unit of the quantum feature extraction module and the key manager of the hierarchical encryption module to realize the full-link adaptive tuning from data preprocessing to model inference.

[0106] In the node priority adjustment stage of the closed-loop iteration, the privacy leakage risk scoring engine analyzes the abnormal event frequency, sensitive data exposure duration, and compliance / non-compliance records in the audit chain, and generates a node trustworthiness index through a weighted scoring algorithm. The trustworthiness index is input into the federated learning scheduler to dynamically calculate the priority weights of the participating nodes, demote the nodes with historical violation records, and at the same time increase the weights of the nodes with high data distribution diversity to optimize the global model convergence efficiency of federated learning.

[0107] The noise variance optimization module receives the real-time model accuracy monitoring data stream, including the decline rate of the F1 value of the validation set and the fluctuation range of the training loss function. Based on the gradient descent algorithm, it searches for the optimal noise variance on the privacy-utility Pareto curve, and combines the geofence parameters in the data cross-border transmission rules of Region A to dynamically adjust the whitelist threshold of the compliance routing engine. The optimized noise variance updates the base value parameters through the configuration interface of the quantum random number generator, and the regional transmission constraints are written into the routing rule database.

[0108] The parameter feedback channel encapsulates the node priority table, the updated noise variance, and the regional constraint parameters into a policy update instruction set, and distributes them to the data preprocessing and encryption module through the control plane. The dimensionality reduction unit of the quantum feature extraction module receives the new regional constraint parameters and dynamically adjusts the weight coefficient of the geographical label in the feature selection algorithm. The key manager of the hierarchical encryption module updates the key rotation policy according to the node priority, enables the enhanced quantum key distribution protocol for the communication links of high-priority nodes, and completes the collaborative optimization of all-link parameters from data input to model output.

[0109] The following explains the technical features in the technical solution of the present invention: Quantum feature extraction: It refers to the technology of reducing the dimensionality of high-dimensional data by using the parallel processing ability of quantum computing. By accelerating the diagonalization of the covariance matrix through the quantum phase estimation algorithm, low-dimensional feature vectors strongly correlated with sensitive attributes (such as user identity, location information) are extracted. Its core lies in the parallel computing characteristics of the superposition state of quantum bits, which significantly improves the computing efficiency compared with classical algorithms.

[0110] Multi-objective optimization hierarchical algorithm: It is a decision-making method that takes data sensitivity, service availability, and regulatory compliance as optimization objectives and generates the Pareto optimal solution set through the NSGA-II algorithm. Data sensitivity is scored based on the correlation between features and sensitive attributes, service availability is based on the contribution of features to model prediction, and compliance matches the regulatory constraints of the target region (such as GDPR). Finally, a privacy level mapping table guiding the encryption policy is generated.

[0111] Dynamic adjustment in the federated learning framework: Based on the Gaussian process model constructed by the Bayesian optimizer, it predicts the aggregation weights of federated nodes; combines the quantum annealing algorithm to map the node selection problem to the Hamiltonian of the Ising model, and filters out a subset of nodes with low latency and complementary data distributions. The dynamic adjustment mechanism balances the sample distribution deviation through weight normalization and the data volume factor, and alleviates the parameter conflicts caused by non-independent and identically distributed (Non-IID) data.

[0112] Quantum Fully Homomorphic Encryption: An encryption scheme based on lattice cryptography that supports direct addition and multiplication operations on ciphertext data. Applied to highly sensitive hierarchical parameters (such as output layer weights), it maintains the encrypted state during the federated aggregation process, avoiding information leakage caused by plaintext transmission, and is also compatible with subsequent noise injection and compression operations.

[0113] Quantum Sparse Compression: Project the encrypted parameters onto a low-dimensional manifold through a quantum variational autoencoder, and use the combination of sparse basis vectors to reduce the volume of ciphertext data. The compression process retains the homomorphic operation ability, reduces communication overhead while maintaining data privacy, and the compressed parameters are transmitted to the downstream module through a quantum secure channel.

[0114] Dynamic Noise Injection Strategy: The deep deterministic policy gradient (DDPG) reinforcement learning model uses the privacy budget consumption rate and model accuracy loss as feedback signals to dynamically generate noise variance and injection position parameters. The quantum random number generator generates a true random noise sequence, which is superimposed on the ciphertext of the encrypted parameters through quantum homomorphic addition to achieve a balance between differential privacy protection and model utility.

[0115] Quantum Hash Comparison: Use the Grover algorithm to construct a quantum circuit, and accelerate the parameter difference localization through the principle of quantum state amplitude amplification. It is used to detect the incrementally updated part of the noisy encrypted parameters, generate a difference matrix and trigger the compression and encryption process, significantly improving the comparison efficiency of large-scale parameter sets.

[0116] Incremental Update and Blockchain Verification: After the differential parameters are sparsely compressed and re-encrypted, the zero-knowledge proof is called through a smart contract to verify the legality of the modification. The verified incremental data is merged into the global model, and the update log is written into the blockchain through a quantum secure hash signature to form an immutable audit chain, supporting version rollback and operation traceability.

[0117] Quantum Anomaly Detection: The quantum support vector machine maps the user request features to a high-dimensional space through a quantum kernel function to identify abnormal behaviors such as high-frequency queries. The dynamic threshold is adjusted according to the historical access pattern and compliance requirements, triggering output fuzzification processing (such as converting probability values to risk interval labels) to block the illegal transmission of sensitive data.

[0118] Audit Chain Generation and Feedback Optimization: Sign the operation log using a quantum secure hash algorithm based on lattice cryptography to generate a blockchain audit chain with a timestamp. Risk metrics (such as the frequency of privacy leakage) are input into the quantum hybrid optimization framework to drive the dynamic tuning of encryption intensity and noise rules, forming a closed-loop feedback link from data transmission, model training to policy iteration.

[0119] Closed-loop Iterative Mechanism: Continuously adjust the encryption key length, noise variance, and node selection weights through a quantum-classical hybrid optimization framework, and feedback the parameters to the data preprocessing module (such as the feature dimensionality reduction weight). The model convergence speed in the audit chain and the risk indicators inversely optimize the full-link strategy to maintain the global optimum of privacy protection and model utility.

[0120] The above technical features collaborate through quantum computing acceleration, dynamic strategy optimization, and hierarchical encryption mechanisms to solve the parameter conflict and privacy leakage problems in the federated learning under non-independent and identically distributed data scenarios, and achieve the full-link privacy-utility balance from data input, model training to inference output.

[0121] Quantum Principal Component Analysis (Quantum PCA): A dimensionality reduction algorithm based on quantum computing, which accelerates the diagonalization of the covariance matrix through quantum phase estimation and controlled rotation gate operations, and extracts the low-dimensional eigenvectors with the strongest correlation with sensitive attributes (such as user identity, location information) in high-dimensional data. Compared with classical PCA, quantum parallelism significantly improves the feature decomposition efficiency and provides efficient feature screening capabilities for subsequent privacy grading.

[0122] Multi-objective Optimization Hierarchical Algorithm (NSGA-II): An improved non-dominated sorting genetic algorithm, which takes the data sensitivity score, business availability weight, and GDPR compliance constraints as optimization objectives, parallelly evaluates the candidate solution sets on the quantum computing unit, and generates the Pareto front solutions covering different privacy-utility balance points. Select the optimal grading strategy through multi-criteria decision analysis, construct the field-level privacy level mapping table, and guide the differential encryption and noise injection rules.

[0123] Bayesian Optimizer: A hyperparameter tuning method based on the Gaussian process model, which fits the historical contribution data of federated learning nodes (such as the parameter update amplitude, convergence stability) through the kernel function, and predicts the aggregation weight allocation ratio in the future training cycle. This optimizer dynamically adjusts the weight coefficients to alleviate the negative impact of non-independent and identically distributed data on the global model convergence.

[0124] Quantum Annealing Algorithm: Map the federated node selection problem to the Ising model Hamiltonian (the node communication delay is mapped to the coupling strength, and the data distribution similarity is mapped to the spin interaction energy), and use the quantum tunneling effect to break through the local optimal solution limit, and quickly screen the node subset with low latency and complementary data distribution. This algorithm optimizes the node combination search efficiency and reduces the parameter deviation caused by heterogeneous data.

[0125] Deep Deterministic Policy Gradient (DDPG) Model: A reinforcement learning model that combines policy gradient and Q-learning. Using the privacy budget consumption rate and model accuracy loss as the joint reward function, it generates the position index and variance intensity policy of noise injection. Dynamically adjusts the noise configuration through the Actor-Critic network structure to balance the differential privacy protection intensity and model usability.

[0126] Quantum Support Vector Machine (Quantum SVM): Utilizes the quantum kernel trick to map user request features to a high-dimensional Hilbert space and constructs a non-linear classification boundary through quantum Fourier transform. Used for real-time detection of high-frequency query abnormal behaviors, combined with dynamic thresholds (based on historical access patterns and compliance requirements) to trigger output fuzzification processing and block potential privacy leakage risks.

[0127] Quantum Approximate Optimization Algorithm (QAOA): A hybrid quantum-classical optimization algorithm that constructs a Hamiltonian containing encryption key length and noise variance parameters, and searches for the privacy-utility Pareto optimal solution through quantum variational optimization loops. This algorithm collaborates with classical gradient descent to dynamically adjust the weight allocation strategy of federated nodes to form a globally optimal parameter combination.

[0128] Lattice-based Cryptography: A quantum-resistant encryption scheme based on lattice mathematical problems, used to construct a quantum fully homomorphic encryption protocol. Generates public and private keys through the Ring-Learning with Errors (RLWE) problem, supports addition and multiplication operations in the ciphertext state, and ensures the computational privacy of highly sensitive parameters during the federated aggregation and noise injection processes.

[0129] Quantum Variational Autoencoder: A quantum-classical hybrid neural network. The encoder quantum circuit learns the sparse representation pattern of encryption parameters and projects high-dimensional ciphertext data into a low-dimensional sparse space. The decoder reconstructs the original data distribution, and the compression process maintains the compatibility of homomorphic operations, significantly reducing the ciphertext transmission and storage overhead.

[0130] Quantum-safe Hash: A quantum-resistant hash function constructed based on lattice cryptography that performs digital signatures on operation logs to generate tamper-proof digest information. Combined with the permissioned-chain blockchain technology, it constructs a timestamped audit chain to support operation traceability and compliance verification, and resist the risk of data forgery under quantum computing attacks.

[0131] Multi-objective Optimization Hierarchical Model (NSGA-II): An improved non-dominated sorting genetic algorithm is adopted, with data sensitivity score, business availability weight, and GDPR compliance constraint as optimization objectives. The candidate solution set is evaluated in parallel on quantum computing units to generate Pareto front solutions covering different privacy-utility balance points. The optimal hierarchical strategy is selected through multi-criteria decision analysis, and a field-level privacy level mapping table is constructed to guide differential encryption and noise injection rules.

[0132] Bayesian Optimization Model (Gaussian Process): A federated node weight prediction model based on Gaussian process regression. It fits the historical contribution data of nodes (such as parameter update amplitude, convergence stability) through kernel functions to predict the aggregation weight allocation ratio in future training cycles. This model dynamically adjusts the weight coefficients to mitigate the negative impact of non-independent and identically distributed data on the convergence of the global model.

[0133] Quantum Annealing Model (Ising Model Mapping): Transforms the federated node selection problem into an Ising model Hamiltonian, where the node communication delay is mapped to the coupling strength coefficient, and the data distribution similarity is calculated through KL divergence and mapped to the spin interaction energy. The quantum tunneling effect of the quantum annealing machine is used to break through the local optimal solution limit and quickly screen a subset of nodes with low latency and complementary data distribution.

[0134] Deep Deterministic Policy Gradient Model (DDPG): A reinforcement learning model that generates noise injection strategies through an Actor-Critic network structure. The Actor network takes the privacy budget consumption rate and model accuracy loss as a joint reward function and outputs noise position and variance parameters; the Critic network evaluates the strategy effect and feedbacks gradient signals to dynamically optimize the privacy-utility balance of noise configuration.

[0135] Quantum-Secure Hash Model (Lattice-based Hash): A quantum-resistant hash function constructed based on lattice cryptography performs digital signatures on operation logs to generate tamper-proof digest information. Combining with permissioned chain blockchain technology, a timestamped audit chain is constructed to support operation traceability and compliance verification, and resist the risk of data forgery under quantum computing attacks.

[0136] Quantum Approximate Optimization Model (QAOA): A hybrid quantum-classical optimization framework constructs a Hamiltonian containing encryption key length and noise variance parameters, and searches for privacy-utility Pareto optimal solutions through quantum variational optimization loops. This model collaborates with classical gradient descent to dynamically adjust the federated node weight allocation strategy to form a globally optimal parameter combination.

[0137] Lattice Cryptography Fully Homomorphic Encryption Model (RLWE-based FHE): A fully homomorphic encryption scheme based on the Ring-Learning with Errors (RLWE) problem that generates a public-private key pair to encrypt highly sensitive parameters and supports addition and multiplication operations in the ciphertext state. This model ensures the privacy of parameter calculations during the federated aggregation process and is also compatible with noise injection and sparse compression operations.

[0138] The present invention relates to a method for model encryption and privacy protection of artificial intelligence algorithms. Aiming at the data heterogeneity and privacy-utility balance problems faced by the federated learning framework in high-sensitive data scenarios, a privacy protection system with full-link collaboration is constructed through the integration of quantum computing and dynamic optimization technologies. The specific implementation is as follows: In the data preprocessing stage, the original business data undergoes feature dimensionality reduction by the quantum principal component analysis module. The quantum phase estimation algorithm is used to accelerate the diagonalization of the covariance matrix, and low-dimensional feature vectors strongly correlated with sensitive attributes are extracted. During the dimensionality reduction process, the feature retention dimension is dynamically determined based on the variance explanation rate, eliminating redundant features while preserving the data distribution pattern. After receiving the dimensionality-reduced features, the multi-objective optimization hierarchical algorithm takes the data sensitivity score, business availability weight, and GDPR compliance metrics as optimization objectives, and generates a Pareto optimal solution set through the improved NSGA-II algorithm to construct a field-level privacy level mapping table. This table divides the data into public, internal, and confidential levels, guiding subsequent hierarchical encryption strategies. For example, confidential-level fields are bound with quantum fully homomorphic encryption and dynamic noise injection mechanisms.

[0139] In the federated learning framework, the aggregation weights of participating nodes are dynamically adjusted through a Gaussian process model constructed by a Bayesian optimizer, combining the node's historical contribution degree and real-time communication delay data to predict the weight coefficients. The quantum annealing machine maps the node selection problem to an Ising model Hamiltonian to solve for a subset of nodes with low latency and complementary data distribution. The dynamic weights and the optimized node subset are input into an improved federated aggregation formula to perform a hierarchical weighted summation operation in the ciphertext state, generating global model parameters. For highly sensitive layer parameters, a quantum fully homomorphic encryption algorithm based on lattice cryptography is applied for ciphertext conversion, and the ciphertext volume is compressed through the sparse coding layer of the quantum neural network to reduce the transmission overhead.

[0140] The noise injection module dynamically generates a noise configuration strategy through a deep deterministic policy gradient reinforcement learning model, using the privacy budget consumption rate and model accuracy loss as feedback signals to optimize the noise injection position and variance intensity. The quantum random number generator generates a true random noise sequence that conforms to a Gaussian distribution, and the noise is superimposed on the encrypted parameter ciphertext using quantum homomorphic addition. The parameters after noise injection are subjected to quantum hash comparison by the Grover algorithm to locate the incrementally updated part and generate a compressed ciphertext increment list, which triggers a global model version iteration after verification by the blockchain smart contract.

[0141] During the user inference stage, the anomaly detection model constructed by the quantum support vector machine analyzes the request features in real time. When the high-frequency queries exceed the dynamic threshold, it triggers the output for fuzzification processing. The probability values are mapped to discrete risk labels through quantum fuzzy logic. Combined with the geographical rules matching of the compliance routing engine, it intercepts the illegal transmission requests. All operation logs are signed by the quantum secure hash algorithm and written into the permissioned blockchain, forming an immutable audit chain. The risk indicators in the audit chain drive the quantum hybrid optimization framework to dynamically adjust the encryption strength and noise rules. For example, it searches for the optimal combination of the key length and noise variance through the quantum approximate optimization algorithm, and feeds back the optimized parameters to the federated node selection and encryption module.

[0142] The closed-loop iterative mechanism continuously monitors the model convergence speed and the risk of privacy leakage, and reversely tunes the feature dimensionality reduction weights and encryption strategies in the data preprocessing stage. For example, the abnormal node behavior data in the audit chain triggers the reallocation of the federated participation priorities, and the model accuracy loss signal drives the dynamic calibration of the noise variance threshold. The full-link parameters are updated collaboratively under the quantum-classical hybrid optimization framework, achieving the global optimum of privacy-utility from data input, model training to inference output, and effectively solving the parameter conflict and privacy leakage risks in the non-independent and identically distributed data scenario.

[0143] The technical solution of the present invention systematically solves the dual contradictions of data heterogeneity and privacy protection in federated learning by constructing a dynamic collaborative optimization system. First, aiming at the parameter conflict caused by non-independent and identically distributed data, it adopts the quantum optimization node selection and multi-objective weighted aggregation strategy. It analyzes the node data distribution similarity matrix through the quantum annealing algorithm, dynamically screens the subset of participating nodes with strong data complementarity, and reduces the divergence of the parameter update direction. At the same time, it optimizes the aggregation weight allocation based on the NSGA-II algorithm, establishes the Pareto optimal weight coefficients under the privacy budget constraint, balances the contribution degree differences of data heterogeneous nodes, and alleviates the model oscillation.

[0144] Second, aiming at the deterioration of convergence stability caused by the privacy protection mechanism, it designs a collaborative mechanism of hierarchical encryption and adaptive noise regulation. It performs quantum fully homomorphic encryption on the highly sensitive parameters according to the privacy level mapping table, and executes the aggregation operation in the ciphertext state to avoid the leakage risk of plaintext transmission. It dynamically adjusts the noise injection intensity and position in combination with the deep deterministic policy gradient model, uses the model accuracy loss as the feedback signal, and adaptively optimizes the noise variance parameters during the training process to achieve the dynamic balance between the privacy protection strength and the model convergence rate.

[0145] Furthermore, the full-link parameter coordination is achieved through an audit-driven closed-loop feedback mechanism. Based on the privacy leakage risk indicators recorded by the quantum security audit chain, the federal node selection strategy and noise injection rules are optimized in reverse; the quantum hybrid optimization framework is used to synchronously adjust the encryption key length and noise distribution parameters, and the optimization results are fed back to the feature dimensionality reduction unit of the data preprocessing module, forming a full-cycle adaptive tuning from data input, model training to inference output, and finally achieving efficient aggregation and global optimality of privacy-utility in a data heterogeneous environment.

Claims

1. A model encryption and privacy protection method for artificial intelligence algorithms, characterized in that: include: Step S101, obtaining original business data, identifying sensitive attributes through quantum feature extraction and dimensionality reduction processing, and generating a dimensionality reduction feature vector; Step S102, inputting the dimension-reduced feature vector into a multi-objective optimization grading algorithm to generate a privacy level mapping table including data sensitivity levels and cleaned structured data; Step S103, based on the privacy level mapping table, dynamically adjust the aggregation weights of participating nodes in the federated learning framework, select a low-latency node subset through quantum optimization, and generate global model parameters; Step S104, performing quantum fully homomorphic encryption on the high-sensitivity layer in the global model parameters according to the privacy level mapping table, and performing quantum sparse compression on the encrypted parameters; Step S105, injecting dynamic noise generated based on reinforcement learning into the compressed encryption parameters to generate noisy encryption model parameters; Step S106, locating the incremental update part of the noisy encryption model parameters by quantum hash comparison, generating an incremental list and updating the global model; Step S107, receiving a user reasoning request, using quantum anomaly detection to identify high-frequency query behaviors, and performing fuzzification processing on the output results of the updated global model; Step S108, intercepting illegal transmissions according to data region label matching compliance rules, and generating an audit chain through quantum secure hashing; Step S109, based on the risk indicators in the audit chain, dynamically adjust the encryption strength and noise injection strategy, and feed back the adjusted strategy parameters to the federated learning node selection and encryption module.

2. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1 is characterized in that: The step S101 includes: The feature covariance matrix is ​​calculated in parallel through quantum principal component analysis to extract the feature vectors associated with sensitive attributes; Input the feature vector into a multi-objective optimization classification algorithm, and use data sensitivity classification, business availability score and GDPR compliance index as optimization targets to generate a Pareto optimal solution set; A privacy level mapping table is generated according to the optimal solution in the Pareto optimal solution set, wherein the privacy level mapping table includes encryption level tags for each data field and is used to control the encryption strength of local training parameters of the federated learning participating nodes.

3. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1, characterized in that: The step S103 includes: A Gaussian process model is built based on the Bayesian optimizer to predict the weight value of each federated learning participating node according to its historical aggregated contribution; Map the node communication delay and data distribution similarity into the Ising model Hamiltonian, and use the quantum annealing machine to solve the optimal federated node subset; According to the predicted weight value of each federated learning participating node and the optimal federated node subset, the weighted average coefficient in the federated aggregation formula is adjusted to generate global model parameters.

4. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1 is characterized in that: The step S104 includes: Encrypting by applying a quantum fully homomorphic encryption algorithm according to the global model parameters marked as a high-sensitivity level in the privacy level mapping table; The encrypted parameters are compressed through the sparse coding layer of the quantum neural network to generate compressed ciphertext parameters; The compressed ciphertext parameters are input into the homomorphic operation interface of the differential privacy noise injection module to maintain the noise superposition in the ciphertext state.

5. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1 is characterized in that: The step S105 comprises: Through the deep deterministic policy gradient reinforcement learning model, the privacy budget consumption rate and model accuracy loss are used as reward functions to generate the model level position and variance intensity strategy for noise injection; Use quantum random number generator to generate noise sequence that conforms to Gaussian distribution; Through the homomorphic addition operation of quantum fully homomorphic encryption, the noise sequence is superimposed on the ciphertext of the encryption parameter to generate the encryption model parameter with noise.

6. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1 is characterized in that: The step S106 includes: Based on the Grover algorithm, the quantum state comparison of the current global model parameters and the historical version is performed to locate the newly added or modified parameters and generate an incremental parameter list; Performing quantum sparse coding processing on the incremental parameter list, and generating compressed ciphertext through a quantum fully homomorphic encryption algorithm; The compressed ciphertext is input into the blockchain storage interface of the audit tracking module, triggering a global model update and generating updated global model parameters.

7. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1 is characterized in that: The step S107 includes: The quantum support vector machine is used to perform quantum kernel mapping on user request features. When the number of requests per unit time exceeds the preset threshold, it is determined to be a high-frequency query abnormal behavior. After the defense mechanism is triggered, the interval boundary value is dynamically adjusted according to the ROC curve of the preset model verification set, and the probability value output by the updated global model is converted into interval labels of high risk, medium risk, and low risk; The abnormal behavior log is input into the rule matching module of the compliance routing engine to block the transmission of illegal data streams including sensitive fields.

8. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1 is characterized in that: The step S108 includes: A quantum-safe hash algorithm based on lattice cryptography is used to digitally sign federated learning node selection records and noise injection operation logs; Synchronize the signed log data to the permissioned blockchain node to generate an unalterable audit chain including a timestamp; The privacy leakage risk index in the audit chain is input into the optimization objective function of the quantum hybrid optimization framework to drive the dynamic adjustment of the encryption strength parameters.

9. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 1 is characterized in that: The step S109 includes: Search for the optimal combination of key length and differential privacy noise variance of quantum fully homomorphic encryption through quantum approximate optimization algorithm; Combined with the classic gradient descent algorithm, the weight distribution strategy for federated node selection is locally optimized to generate updated encryption strength parameters and noise injection rules; The encryption strength parameters and noise injection rules are synchronized in real time to the aggregation controller and differential privacy module of the federated learning framework.

10. The model encryption and privacy protection method for artificial intelligence algorithms according to claim 9, characterized in that: Also includes: According to the privacy leakage risk score in the audit chain, the selection priority of the federated learning participating nodes is adjusted inversely; Based on the real-time monitored model accuracy loss data, the differential privacy noise variance and the regional transmission constraints in the compliance routing rules are optimized; The adjusted node selection priority, noise variance and regional transmission constraint parameters are fed back to the dimension reduction processing unit of the quantum feature extraction module and the key manager of the hierarchical encryption module.

Images