Multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning

Abstract

The present invention provides a multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning, which relates to the technical field of data sharing and privacy protection, including: adding mixed noise to the original visitor portrait data for privacy protection by calculating feature sensitivity and allocating differential privacy budget, encrypting feature vectors by fully homomorphic encryption based on ideal lattice, performing threshold cipher segmentation by using Viterbi algorithm, distributing sub-key components to each private domain environment, constructing a hybrid network structure to extract features, performing federated optimization based on data quality and historical contribution, updating the feature extraction network, performing feature dimensionality reduction by using dual autoencoders, extracting key features and generating user feature fingerprints, grouping users by calculating feature fingerprint similarity, and realizing secure sharing of user similarity information and original visitor portrait data in combination with multi-party secure computing and hierarchical data access strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of data sharing and privacy protection, and in particular to a multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning. Background Art

[0002] With the popularization of the Internet and mobile devices, enterprises have accumulated a large amount of visitor data. How to effectively use visitor data to build a comprehensive and accurate visitor portrait has become an urgent problem to be solved;

[0003] Traditional visitor profile building methods usually rely on centralized data storage and processing, that is, collecting all data to a central server for analysis. However, the centralized approach has problems such as data silos, privacy leakage, and data security.

[0004] Federated learning, as a distributed machine learning paradigm, allows multiple participants to conduct collaborative training without sharing original data, effectively solving the problems of data silos and privacy protection. However, the existing visitor profiling methods based on federated learning still have problems such as imperfect feature sensitivity quantification and privacy protection mechanism, low model aggregation efficiency and accuracy, and inflexible data sharing and access control mechanisms.

[0005] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0006] The embodiment of the present invention provides a multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning, including:

[0008] Obtain original visitor portrait data in multiple independent private domain environments, calculate the information entropy corresponding to each category of data in the original visitor portrait data based on JS divergence, obtain a quantitative indicator of feature sensitivity, calculate the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtain the differential privacy budget allocation result and construct a mixed noise generator, construct a mixed noise sequence and add it to the original visitor portrait data to generate a privacy-preserving feature vector, generate a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypt the privacy-preserving feature vector to obtain an encrypted feature vector, perform threshold cryptographic segmentation on the fully homomorphic encryption key through the Viterbi algorithm, obtain multiple sub-key components, and send the sub-key components to the independent private domain environment through a distributed consistency protocol;

[0009] A hybrid network structure is constructed based on a multi-level recursive neural network and a graph neural network, an adaptive attention mechanism and a residual connection mechanism are added to each layer of the hybrid network structure to obtain a feature extraction network, the encrypted feature vector is added to the feature extraction network to obtain an initial feature representation, a data quality assessment index is determined based on the initial feature representation, the historical contribution of each independent private domain environment is calculated by a time series weighted decay algorithm, the data quality assessment index is weightedly fused with the historical contribution to obtain a model aggregation weight, a two-layer asynchronous federated optimization framework is constructed, and the feature extraction network is updated and anomaly detected based on the model aggregation weight, and effective local model parameters are determined to obtain an updated feature extraction network;

[0010] The updated feature extraction network and encrypted feature vector are added to the dual autoencoder structure to perform feature dimensionality reduction to obtain intermediate feature representation, key features are extracted through the sparse autoencoder layer, and the feature fingerprint corresponding to each user is generated according to the preset feature fingerprint generation function, the similarity between different feature fingerprints is calculated and clustered to obtain the initial feature cluster, and the information loss minimization algorithm is executed to group users, and the user node similarity matrix is ​​calculated based on the multi-layer heterogeneous feature interaction graph and the graph attention network. The user node similarity matrix is ​​partially decrypted based on the sub-key component, and the partial decryption results corresponding to each independent private domain environment are aggregated in combination with the multi-party secure computing protocol to obtain complete user similarity information and assign decryption components to each user according to the pre-built hierarchical data access strategy, which is combined with the data transmission channel built based on the hierarchical dynamic key negotiation mechanism for secure sharing.

[0011] In an optional embodiment,

[0012] Obtaining original visitor portrait data in multiple independent private domain environments, calculating the information entropy corresponding to each category of data in the original visitor portrait data based on JS divergence, obtaining a quantitative indicator of feature sensitivity, calculating the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtaining the differential privacy budget allocation result and constructing a mixed noise generator, constructing a mixed noise sequence and adding it to the original visitor portrait data, generating a privacy-preserving feature vector, generating a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypting the privacy-preserving feature vector to obtain an encrypted feature vector, performing threshold cryptographic segmentation on the fully homomorphic encryption key through a Viterbi algorithm, obtaining multiple sub-key components, and sending the sub-key components to the independent private domain environment through a distributed consistency protocol, including:

[0013] Obtaining original visitor portrait data in multiple independent private domain environments, wherein the original visitor portrait data includes user ID, age, gender, region, browsing history, and purchase history;

[0014] Calculate the probability distribution of each data category in the original data of the visitor portrait and the average distribution of the overall data, substitute the probability distribution of the data category and the average distribution into the Kuller-Leibler divergence formula to calculate a first divergence value, substitute the probability distribution of the overall data and the average distribution into the Kuller-Leibler divergence formula to calculate a second divergence value, perform weighted averaging of the first divergence value and the second divergence value to obtain the Jensen-Shannon divergence, and normalize the Jensen-Shannon divergence to obtain a feature sensitivity quantification index;

[0015] Construct a differential privacy budget allocation model based on dynamic programming, set the total privacy budget as the target constraint, set the feature sensitivity quantification indicator as the optimization weight, construct a state transfer equation, where the state variable of the state transfer equation is the allocated privacy budget, and the decision variable is the privacy budget to be allocated. Iterate and calculate the optimal substructure based on the state transfer equation to obtain the privacy budget allocation value for each data category;

[0016] Constructing a mixed noise generator, the mixed noise generator including a Laplace noise module and a Gaussian noise module, calculating the numerical sensitivity of each of the data categories, substituting the numerical sensitivity and the corresponding privacy budget allocation value into the Laplace distribution function and the Gaussian distribution function respectively, calculating the noise intensity corresponding to each of the data categories based on the Laplace distribution function and the Gaussian distribution function, adding random noise to each of the data categories through the mixed noise generator according to the noise intensity, obtaining a mixed noise sequence, and adding the mixed noise sequence to the original visitor portrait data to obtain a privacy protection feature vector;

[0017] Constructing security parameters of an ideal lattice cryptographic system, generating two large integers as public keys based on the security parameters, randomly selecting a polynomial as a private key, combining the public key and the private key to form the fully homomorphic encryption key pair, selecting a random noise polynomial, and using the public key in the fully homomorphic encryption key pair to perform encryption operation on the privacy protection feature vector to obtain an encrypted feature vector;

[0018] The private key in the fully homomorphic encryption key pair is coefficient decomposed to obtain a coefficient vector, a polynomial interpolation operation is performed on the coefficient vector to obtain multiple coordinate points, the coordinate points are secretly divided by a Viterbi algorithm to obtain multiple sub-key components, and the multiple sub-key components are respectively distributed to each of the independent private domain environments through a distributed consistency protocol.

[0019] In an optional embodiment,

[0020] The coordinate points are secretly divided by the Viterbi algorithm to obtain multiple sub-key components including:

[0021] Randomly select two prime numbers greater than 100 and multiply them to obtain a modulus, calculate the lowest common multiple of the prime numbers minus one to obtain a private key parameter, select a generator with a specified order, the order of the generator is the square of the modulus, and combine the modulus and the generator to obtain a public key;

[0022] Select a random number for each coordinate point, perform modular exponentiation operation on the horizontal coordinate and the vertical coordinate of the coordinate point, the generator and the random number, respectively, to obtain an encrypted coordinate point, the encrypted coordinate point including an encrypted horizontal coordinate and an encrypted vertical coordinate, and construct all the encrypted coordinate points into an encrypted coordinate point matrix;

[0023] Constructing a state transfer equation, wherein the input variable of the state transfer equation is the encrypted coordinate point, and the output variable of the state transfer equation is the encrypted path score, performing dynamic programming calculation on the ciphertext domain, recursively calculating the optimal path score of each state node based on the state transfer equation, and obtaining a global optimal encrypted path by reverse backtracking, wherein the global optimal encrypted path is composed of multiple encrypted coordinate points connected in a transfer order;

[0024] Divide the encrypted coordinate point matrix based on the global optimal encryption path, divide the encrypted coordinate point matrix into multiple encryption subsets, each encryption subset contains the encrypted coordinate points of the corresponding state in the global optimal encryption path, construct a random polynomial, the constant term of the random polynomial is the sum of the encrypted coordinate points in the encryption subset, and the other coefficients of the random polynomial are encrypted random numbers, calculate the values ​​of the random polynomial at different input values, obtain multiple encryption coefficients, and distribute the encryption coefficients as subkey components to different participants;

[0025] Receive the encrypted subkey components returned by the participant, aggregate the encrypted subkey components, decrypt the aggregated results according to the private key parameters to obtain the plaintext subkey components, substitute the plaintext subkey components into the secret sharing recovery function, and restore the complete private key.

[0026] In an optional embodiment,

[0027] A hybrid network structure is constructed based on a multi-level recursive neural network and a graph neural network, an adaptive attention mechanism and a residual connection mechanism are added to each layer of the hybrid network structure to obtain a feature extraction network, the encrypted feature vector is added to the feature extraction network to obtain an initial feature representation, a data quality evaluation index is determined based on the initial feature representation, the historical contribution of each independent private domain environment is calculated by a time series weighted decay algorithm, the data quality evaluation index and the historical contribution are weightedly fused to obtain a model aggregation weight, a two-layer asynchronous federated optimization framework is constructed, and the feature extraction network is updated and anomaly detected based on the model aggregation weight, and effective local model parameters are determined to obtain an updated feature extraction network, including:

[0028] Based on multi-level recursive neural networks and graph neural networks, multiple long short-term memory network layers and multiple graph convolutional network layers are alternately stacked to build a hybrid network structure;

[0029] Calculating the dot product attention weight matrix of each layer of the hybrid network structure, multiplying the attention weight matrix with the input feature to obtain a weighted feature, adding the weighted feature to the input feature to generate a residual enhancement feature, and repeatedly calculating the weighted feature and the residual enhancement feature in each layer of the hybrid network structure to obtain a feature extraction network;

[0030] Receiving an encrypted feature vector, inputting the encrypted feature vector into the hybrid network structure, outputting an initial feature representation, extracting a missing value ratio, an outlier ratio, and data redundancy in the initial feature representation, generating a quality assessment score, and performing a weighted average calculation on the quality assessment score to obtain a comprehensive data quality assessment index;

[0031] Obtain the historical gradient contribution sequence of each private domain environment, generate an attenuation weight sequence based on the time step, perform a weighted sum operation on the historical gradient contribution sequence and the attenuation weight sequence according to the time series weighted attenuation algorithm, calculate the historical contribution corresponding to each private domain environment, perform a weighted fusion operation on the comprehensive data quality assessment index and the historical contribution, and generate a model aggregation weight for each private domain environment;

[0032] Construct a two-layer asynchronous federated optimization framework, wherein the two-layer asynchronous federated optimization framework includes a global optimization layer and a local optimization layer. The global optimization layer collects uploaded local model parameters from each private domain environment, and performs weighted averaging according to the corresponding model aggregation weights to obtain global model parameters. The global model parameters are transmitted from the central server to the local device of each private domain environment through a secure communication protocol. In the local optimization layer, each private domain environment updates the global model parameters according to local data through a gradient descent algorithm to obtain updated local model parameters that adapt to the local data distribution.

[0033] The gradient norm difference value between the updated local model parameters and the global model parameters is calculated, the local model parameters whose gradient norm difference value exceeds a preset threshold are marked as abnormal parameters and filtered, and the normal parameters are retained for parameter update of the feature extraction network to obtain an updated feature extraction network.

[0034] In an optional implementation, the calculation formula corresponding to the historical contribution is as follows:

[0035] ;

[0036] in, c j Represents private domain environment j The historical contribution of T represents the total number of time steps, t represents the time step index, λ 0 represents the initial attenuation coefficient, α j ( t ) indicates the private domain environment j At time step t data quality, β represents the contribution rate adjustment coefficient, δ j ( t ) indicates the private domain environment j At time step t The gradient change rate, γ represents the gradient influence coefficient, g j ( t ) indicates the private domain environment j At time step t The gradient contribution of .

[0037] In an optional embodiment,

[0038] The updated feature extraction network and encrypted feature vector are added to the dual autoencoder structure to perform feature dimension reduction to obtain intermediate feature representation, key features are extracted through the sparse autoencoder layer and feature fingerprints corresponding to each user are generated according to a preset feature fingerprint generation function, the similarity between different feature fingerprints is calculated and clustered to obtain the initial feature cluster, and the information loss minimization algorithm is executed to group users, and the user node similarity matrix is ​​calculated based on the multi-layer heterogeneous feature interaction graph and the graph attention network, and the user node similarity matrix is ​​partially decrypted based on the sub-key component, and the partial decryption results corresponding to each independent private domain environment are aggregated in combination with the multi-party secure computing protocol to obtain complete user similarity information and assign decryption components to each user according to the pre-built hierarchical data access strategy, and the data transmission channel built based on the hierarchical dynamic key negotiation mechanism is combined for secure sharing, including:

[0039] Inputting the updated feature extraction network and the encrypted feature vector into a dual autoencoder structure, the dual autoencoder structure comprising a first autoencoder and a second autoencoder, inputting the encrypted feature vector into the encoder of the first autoencoder, mapping the encrypted feature vector to a low-dimensional space through a multi-layer feedforward neural network to obtain a first compressed feature, inputting the first compressed feature into the decoder of the first autoencoder to reconstruct a preliminary compressed feature representation, inputting the preliminary compressed feature representation into the encoder of the second autoencoder to reduce the feature dimension to obtain a second compressed feature, and inputting the second compressed feature into the decoder of the second autoencoder to reconstruct an intermediate feature representation;

[0040] The intermediate feature representation is input into a sparse autoencoder layer, the activation sparsity of the hidden layer nodes is modified by adding a norm regularization term to the loss function of the sparse autoencoder layer, the loss function is optimized by a back propagation algorithm to obtain a sparse feature representation, a local sensitive hash function is used as a feature fingerprint generation function, the sparse feature representation is mapped into a binary feature fingerprint of a fixed length, for the feature fingerprints of any two users, the number of different binary bits corresponding to each feature fingerprint is counted as the Hamming distance to obtain a user similarity matrix, and the most similar user pairs are iteratively merged based on the user similarity matrix through a bottom-up hierarchical clustering algorithm until a preset number of clusters is reached to obtain an initial feature cluster;

[0041] Construct a multi-layer heterogeneous feature interaction graph, where each user is represented as a node in the graph, and the interaction relationship between users is represented as an edge. The user features are divided into multiple dimensions, and a layer of feature interaction subgraph is constructed for each dimension. In the feature interaction subgraph corresponding to each layer, the weight of the edge is determined based on the feature similarity of the user in the current dimension. The multi-layer heterogeneous feature interaction graph is input into the graph attention network, and the attention coefficient between the central node and the neighboring nodes is calculated. The neighboring node features are weighted summed based on the attention coefficient. The node representation is iteratively updated through the multi-layer graph attention network, and the user node similarity matrix is ​​calculated based on the updated node representation.

[0042] For each private domain environment, partially decrypt the user node similarity matrix according to the pre-assigned sub-key components, aggregate the partial decryption results of each private domain environment through a secure multi-party computing protocol, build a hierarchical data access strategy based on a role-based access control model, divide users into different permission levels, assign a corresponding attribute set to each permission level, generate a decryption component matching the attribute set based on an attribute encryption algorithm, and assign the corresponding decryption component to users at different permission levels;

[0043] A hierarchical dynamic key negotiation mechanism is adopted to build a data transmission channel, and independent key pools are maintained at the network, transmission and application levels respectively. The transmission key is periodically updated between the levels based on the key exchange protocol, and the complete user similarity information and the original visitor portrait data are securely transmitted between different private domain environments through the data transmission channel.

[0044] In an optional embodiment,

[0045] The partial decryption results of each private domain environment are aggregated through the secure multi-party computing protocol, and a hierarchical data access strategy is built based on the role-based access control model to divide users into different permission levels, including:

[0046] Using secret sharing technology to divide the partial decryption result of the private domain environment, dividing the partial decryption result into multiple secret shares, and sending the multiple secret shares to other private domain environments and trusted third-party coordinators except the current private domain environment respectively;

[0047] The trusted third-party coordinator receives the multiple secret shares sent by the private domain environment, and combines the secret shares through a secret share reconstruction algorithm to obtain complete user similarity information;

[0048] Building a role-based access control model based on the user's identity information, behavior history, and security level, the role-based access control model including multiple roles and multiple permission sets, each of the roles corresponding to one permission set;

[0049] A rule engine is used to dynamically assign roles to the user, and the user is assigned to one or more roles, and the final permission of the user is the union of the permission sets corresponding to the assigned roles;

[0050] During the data access process, the user's role and permissions are checked in real time, and based on predefined access control rules, it is determined whether the user has the permission to access specific data objects. If the role to which the user belongs has the corresponding permission, the user is allowed to access the data; otherwise, the access operation is not allowed.

[0051] A second aspect of an embodiment of the present invention provides a multi-private domain visitor portrait sharing and privacy protection routing system based on federated learning, including:

[0052] The first unit is used to obtain the original data of visitor portraits in multiple independent private domain environments, calculate the information entropy corresponding to each category of data in the original data of the visitor portraits based on JS divergence, obtain a quantitative index of feature sensitivity, calculate the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtain the differential privacy budget allocation result and construct a mixed noise generator, construct a mixed noise sequence and add it to the original data of the visitor portraits, generate a privacy-preserving feature vector, generate a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypt the privacy-preserving feature vector to obtain an encrypted feature vector, perform threshold cryptographic segmentation on the fully homomorphic encryption key through a Viterbi algorithm, obtain multiple sub-key components, and send the sub-key components to the independent private domain environment through a distributed consistency protocol;

[0053] The second unit is used to build a hybrid network structure based on a multi-level recursive neural network and a graph neural network, add an adaptive attention mechanism and a residual connection mechanism to each layer of the hybrid network structure to obtain a feature extraction network, add the encrypted feature vector to the feature extraction network to obtain an initial feature representation, determine a data quality assessment index based on the initial feature representation, calculate the historical contribution of each independent private domain environment through a time series weighted decay algorithm, weightedly fuse the data quality assessment index with the historical contribution to obtain a model aggregation weight, build a two-layer asynchronous federated optimization framework, update and detect anomalies on the feature extraction network based on the model aggregation weight, determine effective local model parameters and obtain an updated feature extraction network;

[0054] The third unit is used to add the updated feature extraction network and the encrypted feature vector to the dual autoencoder structure to perform feature dimensionality reduction to obtain an intermediate feature representation, extract key features through a sparse autoencoder layer and generate a feature fingerprint corresponding to each user according to a preset feature fingerprint generation function, calculate the similarity between different feature fingerprints and cluster them to obtain an initial feature cluster, execute an information loss minimization algorithm to group users, calculate a user node similarity matrix based on a multi-layer heterogeneous feature interaction graph and a graph attention network, partially decrypt the user node similarity matrix based on the sub-key component, aggregate the partial decryption results corresponding to each independent private domain environment in combination with a multi-party secure computing protocol, obtain complete user similarity information, and assign a decryption component to each user according to a pre-built hierarchical data access strategy, and securely share the data transmission channel built based on a hierarchical dynamic key negotiation mechanism.

[0055] According to a third aspect of the embodiments of the present invention,

[0056] An electronic device is provided, comprising:

[0057] processor;

[0058] a memory for storing processor-executable instructions;

[0059] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0060] According to a fourth aspect of the embodiments of the present invention,

[0061] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0062] In the present invention, through differential privacy, homomorphic encryption, threshold secret sharing and other technologies, the privacy and security of the original data of visitor portraits are effectively protected, the leakage of sensitive information is prevented, and the data is available but invisible. By utilizing the federated learning framework and asynchronous optimization strategy, the sharing of visitor portraits among multiple private domain environments is realized without directly exchanging the original data, thereby improving data utilization and avoiding the problem of data islands. Based on hybrid network structure, adaptive attention mechanism and multi-layer heterogeneous feature interaction graph and other technologies, more comprehensive and accurate user features are extracted, user grouping and accurate portrait are realized, and strong support is provided for applications such as personalized recommendation and precision marketing. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A schematic diagram of a process of a multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning in an embodiment of the present invention;

[0064] Figure 2 This is a structural diagram of a multi-private domain visitor portrait sharing and privacy protection routing system based on federated learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0067] Figure 1 Schematic diagram of the process of the multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0068] S1. Obtain the original data of visitor portraits in multiple independent private domain environments, calculate the information entropy corresponding to each category of data in the original data of the visitor portraits based on the JS divergence, obtain the quantitative index of feature sensitivity, calculate the optimal allocation plan of the differential privacy budget through the dynamic programming algorithm, obtain the differential privacy budget allocation result and construct a mixed noise generator, construct a mixed noise sequence and add it to the original data of the visitor portraits, generate a privacy-preserving feature vector, generate a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypt the privacy-preserving feature vector to obtain an encrypted feature vector, perform threshold cryptographic segmentation on the fully homomorphic encryption key through the Viterbi algorithm, obtain multiple sub-key components, and send the sub-key components to the independent private domain environment through a distributed consistency protocol;

[0069] The independent private domain environment refers to a data processing or computing process in which sensitive information is independently stored and processed without external sharing to ensure the privacy and security of the data. The JS divergence is a symmetry measure for measuring the difference between two probability distributions. It is based on the KL divergence but has symmetry and its value is limited to the range of [0, 1], thereby providing a more stable similarity measure. The feature sensitivity quantification index is an indicator that measures the sensitivity of a system or model to changes in input features, and is used to determine which features have the greatest influence in prediction or decision-making. The dynamic programming algorithm is an algorithm that breaks down a problem into sub-problems and solves each sub-problem once, storing the solutions to the sub-problems to avoid repeated calculations. It is widely used in optimal substructure problems. The differential privacy budget refers to the privacy resources allocated to control data privacy leakage under the differential privacy framework, which defines the privacy leakage that can be tolerated during data analysis or query. The mixed noise generator is a mechanism for generating noise from multiple sources or noise distributions to enhance privacy protection or provide more randomness in data processing and encryption. The ideal lattice-based public key encryption algorithm is an encryption system that relies on an ideal lattice structure and uses an ideal lattice in mathematical lattice theory to provide encryption security, especially with strong resistance to quantum computing attacks. The fully homomorphic encryption key pair is a key pair for fully homomorphic encryption. Fully homomorphic encryption allows operations to be performed on encrypted data without first decrypting it, thereby performing calculations while maintaining privacy. The Viterbi algorithm is a dynamic programming algorithm for decoding the most likely hidden state sequence under a given observation sequence, and is often used in hidden Markov models. The threshold cipher splitting is an encryption technique that aims to enhance the security of the key by splitting the key into multiple parts and setting a threshold. The original key can only be reconstructed when the threshold is reached or exceeded.

[0070] In an optional embodiment,

[0071] Obtaining original visitor portrait data in multiple independent private domain environments, calculating the information entropy corresponding to each category of data in the original visitor portrait data based on JS divergence, obtaining a quantitative indicator of feature sensitivity, calculating the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtaining the differential privacy budget allocation result and constructing a mixed noise generator, constructing a mixed noise sequence and adding it to the original visitor portrait data, generating a privacy-preserving feature vector, generating a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypting the privacy-preserving feature vector to obtain an encrypted feature vector, performing threshold cryptographic segmentation on the fully homomorphic encryption key through a Viterbi algorithm, obtaining multiple sub-key components, and sending the sub-key components to the independent private domain environment through a distributed consistency protocol, including:

[0072] Obtaining original visitor portrait data in multiple independent private domain environments, wherein the original visitor portrait data includes user ID, age, gender, region, browsing history, and purchase history;

[0073] Calculate the probability distribution of each data category in the original data of the visitor portrait and the average distribution of the overall data, substitute the probability distribution of the data category and the average distribution into the Kuller-Leibler divergence formula to calculate a first divergence value, substitute the probability distribution of the overall data and the average distribution into the Kuller-Leibler divergence formula to calculate a second divergence value, perform weighted averaging of the first divergence value and the second divergence value to obtain the Jensen-Shannon divergence, and normalize the Jensen-Shannon divergence to obtain a feature sensitivity quantification index;

[0074] Construct a differential privacy budget allocation model based on dynamic programming, set the total privacy budget as the target constraint, set the feature sensitivity quantification indicator as the optimization weight, construct a state transfer equation, where the state variable of the state transfer equation is the allocated privacy budget, and the decision variable is the privacy budget to be allocated. Iterate and calculate the optimal substructure based on the state transfer equation to obtain the privacy budget allocation value for each data category;

[0075] Constructing a mixed noise generator, the mixed noise generator including a Laplace noise module and a Gaussian noise module, calculating the numerical sensitivity of each of the data categories, substituting the numerical sensitivity and the corresponding privacy budget allocation value into the Laplace distribution function and the Gaussian distribution function respectively, calculating the noise intensity corresponding to each of the data categories based on the Laplace distribution function and the Gaussian distribution function, adding random noise to each of the data categories through the mixed noise generator according to the noise intensity, obtaining a mixed noise sequence, and adding the mixed noise sequence to the original visitor portrait data to obtain a privacy protection feature vector;

[0076] Constructing security parameters of an ideal lattice cryptographic system, generating two large integers as public keys based on the security parameters, randomly selecting a polynomial as a private key, combining the public key and the private key to form the fully homomorphic encryption key pair, selecting a random noise polynomial, and using the public key in the fully homomorphic encryption key pair to perform encryption operation on the privacy protection feature vector to obtain an encrypted feature vector;

[0077] The private key in the fully homomorphic encryption key pair is coefficient decomposed to obtain a coefficient vector, a polynomial interpolation operation is performed on the coefficient vector to obtain multiple coordinate points, the coordinate points are secretly divided by a Viterbi algorithm to obtain multiple sub-key components, and the multiple sub-key components are respectively distributed to each of the independent private domain environments through a distributed consistency protocol.

[0078] The Kuhler-Leibler divergence formula is a formula for measuring the difference between two probability distributions. The difference is calculated by considering the cross information of each distribution and has good mathematical properties. The optimal substructure refers to the property that the optimal solution to a problem can be composed of the optimal solutions of its subproblems. This is why the dynamic programming algorithm can efficiently solve many problems. The polynomial interpolation operation is a mathematical method for constructing a polynomial through given discrete data points so that the polynomial passes through all these points, thereby enabling interpolation calculations to be performed between the data points.

[0079] Get the original visitor profile data in multiple independent private environments. Assume there are three independent private environments A, B, and C, and collect the original visitor profile data of users respectively. This data includes user ID, age, gender, region, browsing history, and purchase history. For example, the data of user 1 in private environment A is {ID1, 25 years old, male, Beijing, browsed books and electronic products, and purchased books}.

[0080] Calculate the feature sensitivity quantification index, calculate the probability distribution of each data category and the average distribution of the overall data. For example, the proportion of users between the ages of 20-30 is 40%, and the overall average distribution is the average of the proportion of users in each age group. Substitute the probability distribution of each data category and the average distribution of the overall data into the Kuller-Leibler divergence formula to calculate the first divergence value and the second divergence value. For example, the first divergence value of age is 0.2, and the second divergence value is 0.1. The first divergence value and the second divergence value are weighted averaged to obtain the Jensen-Shannon divergence. For example, the Jensen-Shannon divergence of age is 0.15. Normalize the Jensen-Shannon divergence to obtain the feature sensitivity quantification index. For example, the feature sensitivity quantification index of age is 0.3.

[0081] Allocate differential privacy budget, build a differential privacy budget allocation model based on dynamic programming, set the total privacy budget to 1.0, set the feature sensitivity quantification index to the optimization weight, and build a state transition equation, where the state variable is the allocated privacy budget and the decision variable is the privacy budget to be allocated. For example, if the feature sensitivity quantification index of age is 0.3, the more privacy budget allocated to age, the better the overall privacy protection effect. Iterate and calculate the optimal substructure based on the state transition equation to obtain the privacy budget allocation value for each data category. For example, the privacy budget allocation value for age is 0.3, the privacy budget allocation value for gender is 0.2, the privacy budget allocation value for region is 0.15, and so on.

[0082] Generate a privacy-preserving feature vector and construct a mixed noise generator, including a Laplace noise module and a Gaussian noise module. Calculate the numerical sensitivity of each data category. For example, the numerical sensitivity of age is 1, because the age range is usually within 1 year. Substitute the numerical sensitivity and the corresponding privacy budget allocation value into the Laplace distribution function and the Gaussian distribution function, respectively, and calculate the noise intensity corresponding to each data category. For example, the noise intensity of age is 0.5. Add random noise to each data category through the mixed noise generator according to the noise intensity to obtain a mixed noise sequence. Add the mixed noise sequence to the original visitor portrait data to obtain a privacy-preserving feature vector. For example, the age of user 1 becomes 25.5 years old after adding noise.

[0083] Encrypt the privacy-preserving feature vector, construct the security parameters of the ideal lattice cryptographic system, generate two large integers as public keys based on the security parameters, randomly select a polynomial as the private key, and combine the public key and the private key to form a fully homomorphic encryption key pair. Select a random noise polynomial, use the public key in the fully homomorphic encryption key pair to encrypt the privacy-preserving feature vector, and obtain the encrypted feature vector. For example, the encrypted feature vector of user 1 is a set of ciphertext.

[0084] Distribute the subkey components and perform coefficient decomposition on the private key in the fully homomorphic encryption key pair to obtain a coefficient vector. Perform polynomial interpolation on the coefficient vector to obtain multiple coordinate points. Secretly split the coordinate points using the Viterbi algorithm to obtain multiple subkey components. Distribute the multiple subkey components to each independent private domain environment using a distributed consensus protocol. For example, the private key is divided into three subkey components and sent to private domain environments A, B, and C respectively.

[0085] In this embodiment, differential privacy and fully homomorphic encryption technology are used to effectively protect user sensitive data, prevent privacy leakage, and enhance user trust. While protecting privacy, it supports calculation and analysis of encrypted data to maximize the utilization of data value, promote data sharing and collaborative analysis, and adopt threshold cryptographic splitting and distributed consistency protocols to avoid the risk of single-point key leakage, improve the security of key management, and reduce key management costs.

[0086] In an optional embodiment,

[0087] The coordinate points are secretly divided by the Viterbi algorithm to obtain multiple sub-key components including:

[0088] Randomly select two prime numbers greater than 100 and multiply them to obtain a modulus, calculate the lowest common multiple of the prime numbers minus one to obtain a private key parameter, select a generator with a specified order, the order of the generator is the square of the modulus, and combine the modulus and the generator to obtain a public key;

[0089] Select a random number for each coordinate point, perform modular exponentiation operation on the horizontal coordinate and the vertical coordinate of the coordinate point, the generator and the random number, respectively, to obtain an encrypted coordinate point, the encrypted coordinate point including an encrypted horizontal coordinate and an encrypted vertical coordinate, and construct all the encrypted coordinate points into an encrypted coordinate point matrix;

[0090] Constructing a state transfer equation, wherein the input variable of the state transfer equation is the encrypted coordinate point, and the output variable of the state transfer equation is the encrypted path score, performing dynamic programming calculation on the ciphertext domain, recursively calculating the optimal path score of each state node based on the state transfer equation, and obtaining a global optimal encrypted path by reverse backtracking, wherein the global optimal encrypted path is composed of multiple encrypted coordinate points connected in a transfer order;

[0091] Divide the encrypted coordinate point matrix based on the global optimal encryption path, divide the encrypted coordinate point matrix into multiple encryption subsets, each encryption subset contains the encrypted coordinate points of the corresponding state in the global optimal encryption path, construct a random polynomial, the constant term of the random polynomial is the sum of the encrypted coordinate points in the encryption subset, and the other coefficients of the random polynomial are encrypted random numbers, calculate the values ​​of the random polynomial at different input values, obtain multiple encryption coefficients, and distribute the encryption coefficients as subkey components to different participants;

[0092] Receive the encrypted subkey components returned by the participant, aggregate the encrypted subkey components, decrypt the aggregated results according to the private key parameters to obtain the plaintext subkey components, substitute the plaintext subkey components into the secret sharing recovery function, and restore the complete private key.

[0093] The generator refers to the basic element in group theory that generates all other elements of the entire group through a certain element of the group. The modulus refers to the divisor in integer division. It plays the role of determining the remainder in the modular operation and is the core parameter of the modular operation. The modular exponentiation operation is an operation that calculates integer powers and takes the modulus. It is used in fields such as encryption and digital signatures. It can effectively calculate the exponential operation of large numbers and prevent overflow. The encryption path score refers to the score obtained through a certain evaluation mechanism in the encryption path, which is usually used to measure the security or effectiveness of the path. The plaintext subkey component refers to the sub-part extracted from the encryption key in the encryption algorithm. It is used to gradually encrypt the plaintext during the encryption process.

[0094] Select two prime numbers greater than 100, such as 101 and 103. Multiply these two prime numbers to get the modulus 10403. Calculate the values ​​of these two prime numbers minus one, that is, 100 and 102, and find their lowest common multiple, 5100, as the private key parameter. Select a generator, such as 2, whose order is the square of the modulus, that is, 10403 squared. Combine the modulus 10403 and the generator 2 to form the public key.

[0095] Select a random number for each coordinate point. Suppose there is a coordinate point (3, 5), select a random number 7 for it. Perform a modular exponentiation operation on the horizontal coordinate 3 and the generator 2, that is, calculate the remainder of 2 to the power of 3 divided by 10403. Perform a modular exponentiation operation on the vertical coordinate 5 and the random number 7, that is, calculate the remainder of 7 to the power of 5 divided by 10403. Get the encrypted coordinate point, including the encrypted horizontal and vertical coordinates. Perform such encryption operations on all coordinate points, and form all encrypted coordinate points into a matrix, called the encrypted coordinate point matrix.

[0096] Construct a state transfer equation, the input variable of which is the encrypted coordinate point, and the output variable is the encrypted path score. For example, assuming that the encrypted coordinate point of the current state is (8, 341) and the encrypted coordinate point of the next state is (64, 1024), then calculate an encrypted path score based on these two encrypted coordinate points. Perform dynamic programming calculations on the ciphertext domain. Based on the state transfer equation, recursively calculate the optimal path score of each state node. By backtracking, the global optimal encrypted path is obtained, which is composed of multiple encrypted coordinate points connected in the transfer order.

[0097] Divide the encrypted coordinate point matrix based on the global optimal encryption path, and divide the encrypted coordinate point matrix into multiple encrypted subsets, each of which contains the encrypted coordinate points of the corresponding state in the global optimal path. For example, suppose an encrypted subset contains three encrypted coordinate points (8, 341), (64, 1024) and (512, 4096). Construct a random polynomial, the constant term of the polynomial is the sum of the encrypted coordinate points in the encrypted subset, and the other coefficients are encrypted random numbers. For example, a quadratic polynomial can be constructed, whose constant term is (8+341+64+1024+512+4096) and the other coefficients are randomly generated encrypted numbers. Calculate the value of the polynomial at different input values ​​(such as 1, 2, 3) to obtain multiple encrypted coefficients. Distribute these encrypted coefficients as subkey components to different participants.

[0098] Receive the encrypted subkey components returned by the participants, aggregate them, decrypt the aggregated results using the private key parameter 5100, obtain the plaintext subkey components, substitute the plaintext subkey components into the secret sharing recovery function, and restore the complete private key.

[0099] In this embodiment, the coordinate points and private keys are protected by the Viterbi algorithm and secret splitting in the ciphertext domain, which effectively prevents information leakage and attacks. The sub-key components are distributed to different participants, which increases the flexibility of key management and avoids the risk of single point failure. The use of dynamic programming calculation and secret sharing technology improves the efficiency of encryption and decryption while ensuring security.

[0100] S2. Construct a hybrid network structure based on a multi-level recursive neural network and a graph neural network, add an adaptive attention mechanism and a residual connection mechanism to each layer of the hybrid network structure to obtain a feature extraction network, add the encrypted feature vector to the feature extraction network to obtain an initial feature representation, determine a data quality assessment index based on the initial feature representation, calculate the historical contribution of each independent private domain environment through a time series weighted decay algorithm, perform weighted fusion of the data quality assessment index and the historical contribution to obtain a model aggregation weight, construct a two-layer asynchronous federated optimization framework, and update and detect anomalies on the feature extraction network based on the model aggregation weight, determine effective local model parameters, and obtain an updated feature extraction network;

[0101] The multi-level recursive neural network is a deep learning model that recursively processes the input data at multiple levels and generates complex representations at each layer to capture the hierarchical features of the data. It is usually used to process tasks with temporal dependencies, such as natural language processing and time series prediction. The graph neural network is a type of neural network that learns based on graph structured data. It captures the relationship between nodes and edges in the graph by passing information between edges between nodes. It is widely used in social networks, recommendation systems, drug discovery and other fields. The hybrid network structure refers to a structure that combines multiple different types of neural network models (such as convolutional neural networks and recursive neural networks). The expressiveness and adaptability of the model are enhanced through multimodal data or multi-task learning. The time series weighted decay algorithm is an algorithm that adjusts weights according to the time dependency of time series data. It highlights recent data by attenuating the influence of historical data, thereby better predicting future trends. The two-layer asynchronous federated optimization framework is a distributed learning framework that aims to optimize the model in parallel on multiple devices or nodes while ensuring the data privacy of each node. The asynchronous update and two-layer optimization structure improve the computing efficiency and optimization effect.

[0102] In an optional embodiment,

[0103] A hybrid network structure is constructed based on a multi-level recursive neural network and a graph neural network, an adaptive attention mechanism and a residual connection mechanism are added to each layer of the hybrid network structure to obtain a feature extraction network, the encrypted feature vector is added to the feature extraction network to obtain an initial feature representation, a data quality evaluation index is determined based on the initial feature representation, the historical contribution of each independent private domain environment is calculated by a time series weighted decay algorithm, the data quality evaluation index and the historical contribution are weightedly fused to obtain a model aggregation weight, a two-layer asynchronous federated optimization framework is constructed, and the feature extraction network is updated and anomaly detected based on the model aggregation weight, and effective local model parameters are determined to obtain an updated feature extraction network, including:

[0104] Based on multi-level recursive neural networks and graph neural networks, multiple long short-term memory network layers and multiple graph convolutional network layers are alternately stacked to build a hybrid network structure;

[0105] Calculating the dot product attention weight matrix of each layer of the hybrid network structure, multiplying the attention weight matrix with the input feature to obtain a weighted feature, adding the weighted feature to the input feature to generate a residual enhancement feature, and repeatedly calculating the weighted feature and the residual enhancement feature in each layer of the hybrid network structure to obtain a feature extraction network;

[0106] Receiving an encrypted feature vector, inputting the encrypted feature vector into the hybrid network structure, outputting an initial feature representation, extracting a missing value ratio, an outlier ratio, and data redundancy in the initial feature representation, generating a quality assessment score, and performing a weighted average calculation on the quality assessment score to obtain a comprehensive data quality assessment index;

[0107] Obtain the historical gradient contribution sequence of each private domain environment, generate an attenuation weight sequence based on the time step, perform a weighted sum operation on the historical gradient contribution sequence and the attenuation weight sequence according to the time series weighted attenuation algorithm, calculate the historical contribution corresponding to each private domain environment, perform a weighted fusion operation on the comprehensive data quality assessment index and the historical contribution, and generate a model aggregation weight for each private domain environment;

[0108] Construct a two-layer asynchronous federated optimization framework, wherein the two-layer asynchronous federated optimization framework includes a global optimization layer and a local optimization layer. The global optimization layer collects uploaded local model parameters from each private domain environment, and performs weighted averaging according to the corresponding model aggregation weights to obtain global model parameters. The global model parameters are transmitted from the central server to the local device of each private domain environment through a secure communication protocol. In the local optimization layer, each private domain environment updates the global model parameters according to local data through a gradient descent algorithm to obtain updated local model parameters that adapt to the local data distribution.

[0109] The gradient norm difference value between the updated local model parameters and the global model parameters is calculated, the local model parameters whose gradient norm difference value exceeds a preset threshold are marked as abnormal parameters and filtered, and the normal parameters are retained for parameter update of the feature extraction network to obtain an updated feature extraction network.

[0110] The decaying weight sequence means that in time series analysis or optimization problems, lower weights are given to old data or old information over time, so that the model or algorithm pays more attention to the most recent input data. The gradient norm difference value is a measure of gradient changes in an optimization algorithm, which is used to evaluate the stability of parameter updates during model training, and is usually used to determine whether it is necessary to adjust the learning rate or terminate training.

[0111] Construct a hybrid network structure. Alternately stack multiple LSTM layers and multiple graph convolutional network layers. For example, use three LSTM layers and three graph convolutional network layers, each LSTM layer receives the output of the previous layer as input and outputs sequence features. Each graph convolutional network layer receives the output of the previous layer and the graph structure information represented by the predefined adjacency matrix, and outputs graph features. These two types of network layers are alternately connected, and the output of the final graph convolutional network layer is used as the output of the hybrid network structure.

[0112] Adaptive attention mechanism and residual connection are added to each layer of the hybrid network structure. For each layer, the dot product attention weight matrix is ​​first calculated. For example, the input features are linearly transformed, and then the dot products between the features are calculated, and then normalized by the Softmax function to obtain the attention weight matrix. The attention weight matrix is ​​multiplied by the input features to obtain weighted features. The weighted features are added to the original input features to generate residual enhanced features. This process is repeated at each layer, and finally a feature extraction network is obtained.

[0113] Input the encrypted feature vector into the feature extraction network to obtain the initial feature representation. Assuming that the encrypted feature vector is a 128-dimensional vector, after passing through the feature extraction network, an initial feature representation of 256 dimensions is obtained. Indicators such as the missing value ratio, outlier ratio, and data redundancy are extracted from the initial feature representation. For example, the missing value ratio is 0.1, the outlier ratio is 0.05, and the data redundancy is 0.2. Multiply these indicators by preset weights, such as 0.5, 0.3, and 0.2, and then sum them to obtain a comprehensive data quality assessment indicator, such as 0.1×0.5+0.05×0.3+0.2×0.2=0.105.

[0114] Obtain the historical gradient contribution sequence of each private environment. For example, suppose there are three private environments, and the length of the historical gradient contribution sequence of each environment is 5, namely [0.1, 0.2, 0.3, 0.4, 0.5], [0.2, 0.3, 0.4, 0.5, 0.6], [0.3, 0.4, 0.5, 0.6, 0.7]. Generate a decay weight sequence based on the time step, such as [0.1, 0.2, 0.3, 0.4, 0.5]. According to the time series weighted decay algorithm, perform a weighted sum operation on the historical gradient contribution sequence and the decay weight sequence to obtain the historical contribution of each private environment. For example, the historical contribution of the first private environment is 0.1×0.1+0.2×0.2+0.3×0.3+0.4×0.4+0.5×0.5=0.55. Perform a weighted fusion of the comprehensive data quality assessment index and the historical contribution to generate the model aggregation weight of each private environment. For example, the model aggregation weight of the first private domain environment is 0.105×0.5+0.55×0.5=0.33.

[0115] Construct a two-layer asynchronous federated optimization framework. The global optimization layer collects the local model parameters uploaded by each private domain environment, and performs weighted averaging according to the corresponding model aggregation weights to obtain the global model parameters. The global model parameters are transmitted from the central server to the local device of each private domain environment through a secure communication protocol. In the local optimization layer, each private domain environment updates the global model parameters based on local data through a gradient descent algorithm to obtain updated local model parameters that adapt to the local data distribution.

[0116] Calculate the gradient norm difference between the updated local model parameters and the global model parameters. Mark the local model parameters whose difference value exceeds the preset threshold as abnormal parameters and filter them, retain the normal parameters for parameter update of the feature extraction network, and obtain the updated feature extraction network. For example, assuming the threshold is 0.1, the gradient norm difference value of the first private domain environment is 0.05, and the gradient norm difference value of the second private domain environment is 0.15, then the parameters of the second private domain environment are marked as abnormal and filtered.

[0117] In this embodiment, through data quality assessment and abnormal parameter filtering, the impact of low-quality data and abnormal local models on the global model is effectively reduced, and the robustness of the model is improved. The hybrid network structure combines the advantages of recursive neural networks and graph neural networks, and can better capture complex patterns in the data, thereby enhancing the generalization ability of the model. The federated optimization framework enables model training to be performed locally in each private domain environment without uploading the original data to the central server, effectively protecting data privacy.

[0118] In an optional implementation, the calculation formula corresponding to the historical contribution is as follows:

[0119] ;

[0120] in, c j Represents private domain environment j The historical contribution of T represents the total number of time steps, t represents the time step index, λ 0 represents the initial attenuation coefficient, α j ( t ) indicates the private domain environment j At time step t data quality, β represents the contribution rate adjustment coefficient, δ j ( t ) indicates the private domain environment j At time step t The gradient change rate, γ represents the gradient influence coefficient, g j ( t ) indicates the private domain environment j At time step t The gradient contribution of .

[0121] In this embodiment, factors such as data quality, time decay and gradient contribution are comprehensively considered, which can more comprehensively evaluate the contribution of the private domain environment to the overall learning and avoid the limitations of a single indicator. By adjusting the decay coefficient, contribution rate adjustment coefficient and gradient influence coefficient, the influence weights of time decay, data quality and gradient contribution on the historical contribution can be dynamically adjusted to adapt to different application scenarios and encourage the private domain environment to provide high-quality data and gradient contributions, thereby improving the efficiency and effectiveness of overall learning.

[0122] S3. Add the updated feature extraction network and encrypted feature vector to the dual autoencoder structure to perform feature dimensionality reduction to obtain intermediate feature representation, extract key features through the sparse autoencoder layer and generate feature fingerprints corresponding to each user according to the pre-set feature fingerprint generation function, calculate the similarity between different feature fingerprints and cluster them to obtain the initial feature cluster, execute the information loss minimization algorithm to group users, calculate the user node similarity matrix based on the multi-layer heterogeneous feature interaction graph and graph attention network, partially decrypt the user node similarity matrix based on the sub-key component, aggregate the partial decryption results corresponding to each independent private domain environment in combination with the multi-party secure computing protocol, obtain complete user similarity information and assign decryption components to each user according to the pre-built hierarchical data access strategy, and share securely in combination with the data transmission channel built based on the hierarchical dynamic key negotiation mechanism.

[0123] The feature fingerprint refers to a "fingerprint" representation for uniquely identifying or identifying data by extracting certain unique features of data or signals. It is often used in identity authentication, data tracking and anti-counterfeiting technology. The information loss minimization algorithm is an optimization algorithm whose goal is to minimize the information loss caused by feature selection, data compression or model simplification during data processing or model training, so as to maintain the core information and features of the data. The user node similarity matrix is ​​a matrix constructed by calculating the similarity between users (such as user behavior analysis based on collaborative filtering). It is used to describe the similarity between different users and is widely used in recommendation systems and social network analysis. The sub-key component refers to the decomposition of the key into several parts or groups in the encryption algorithm. Each component is called a sub-key component, which plays a role in different stages of the encryption process to ensure the security and complexity of the encryption process. The multi-party secure computing protocol refers to the use of a security protocol to ensure that the data privacy of each party is not leaked when sensitive data is shared and processed among multiple parties. It is often used in distributed computing and privacy protection scenarios, such as blockchain-based contract execution. The decryption component refers to a module in the encryption system, which is used to restore encrypted data to original plaintext data, and realizes two-way operations of encryption and decryption by working with the encryption component. The hierarchical dynamic key negotiation mechanism is a multi-stage key exchange protocol that allows participants to dynamically generate keys according to the current network status at different communication stages, and improves the security and flexibility of the protocol in a hierarchical manner.

[0124] In an optional embodiment,

[0125] The updated feature extraction network and encrypted feature vector are added to the dual autoencoder structure to perform feature dimension reduction to obtain intermediate feature representation, key features are extracted through the sparse autoencoder layer and feature fingerprints corresponding to each user are generated according to a preset feature fingerprint generation function, the similarity between different feature fingerprints is calculated and clustered to obtain the initial feature cluster, and the information loss minimization algorithm is executed to group users, and the user node similarity matrix is ​​calculated based on the multi-layer heterogeneous feature interaction graph and the graph attention network, and the user node similarity matrix is ​​partially decrypted based on the sub-key component, and the partial decryption results corresponding to each independent private domain environment are aggregated in combination with the multi-party secure computing protocol to obtain complete user similarity information and assign decryption components to each user according to the pre-built hierarchical data access strategy, and the data transmission channel built based on the hierarchical dynamic key negotiation mechanism is combined for secure sharing, including:

[0126] Inputting the updated feature extraction network and the encrypted feature vector into a dual autoencoder structure, the dual autoencoder structure comprising a first autoencoder and a second autoencoder, inputting the encrypted feature vector into the encoder of the first autoencoder, mapping the encrypted feature vector to a low-dimensional space through a multi-layer feedforward neural network to obtain a first compressed feature, inputting the first compressed feature into the decoder of the first autoencoder to reconstruct a preliminary compressed feature representation, inputting the preliminary compressed feature representation into the encoder of the second autoencoder to reduce the feature dimension to obtain a second compressed feature, and inputting the second compressed feature into the decoder of the second autoencoder to reconstruct an intermediate feature representation;

[0127] The intermediate feature representation is input into a sparse autoencoder layer, the activation sparsity of the hidden layer nodes is modified by adding a norm regularization term to the loss function of the sparse autoencoder layer, the loss function is optimized by a back propagation algorithm to obtain a sparse feature representation, a local sensitive hash function is used as a feature fingerprint generation function, the sparse feature representation is mapped into a binary feature fingerprint of a fixed length, for the feature fingerprints of any two users, the number of different binary bits corresponding to each feature fingerprint is counted as the Hamming distance to obtain a user similarity matrix, and the most similar user pairs are iteratively merged based on the user similarity matrix through a bottom-up hierarchical clustering algorithm until a preset number of clusters is reached to obtain an initial feature cluster;

[0128] Construct a multi-layer heterogeneous feature interaction graph, where each user is represented as a node in the graph, and the interaction relationship between users is represented as an edge. The user features are divided into multiple dimensions, and a layer of feature interaction subgraph is constructed for each dimension. In the feature interaction subgraph corresponding to each layer, the weight of the edge is determined based on the feature similarity of the user in the current dimension. The multi-layer heterogeneous feature interaction graph is input into the graph attention network, and the attention coefficient between the central node and the neighboring nodes is calculated. The neighboring node features are weighted summed based on the attention coefficient. The node representation is iteratively updated through the multi-layer graph attention network, and the user node similarity matrix is ​​calculated based on the updated node representation.

[0129] For each private domain environment, partially decrypt the user node similarity matrix according to the pre-assigned sub-key components, aggregate the partial decryption results of each private domain environment through a secure multi-party computing protocol, build a hierarchical data access strategy based on a role-based access control model, divide users into different permission levels, assign a corresponding attribute set to each permission level, generate a decryption component matching the attribute set based on an attribute encryption algorithm, and assign the corresponding decryption component to users at different permission levels;

[0130] A hierarchical dynamic key negotiation mechanism is adopted to build a data transmission channel, and independent key pools are maintained at the network, transmission and application levels respectively. The transmission key is periodically updated between the levels based on the key exchange protocol, and the complete user similarity information and the original visitor portrait data are securely transmitted between different private domain environments through the data transmission channel.

[0131] The multi-layer feedforward neural network is a common neural network architecture, which includes multiple levels of neuron nodes. The output of each layer is used as the input of the next layer, and is usually used for tasks such as classification and regression. The norm regularization term refers to adding a regularization term to the loss function of a machine learning or deep learning model, which is usually the norm of the model parameters (such as the L1 norm or the L2 norm) to control the complexity of the model and prevent overfitting. The local sensitive hash function is a data mapping technology used to map high-dimensional data to a low-dimensional space to ensure that similar data still maintains a high similarity after mapping. It is widely used in tasks such as approximate nearest neighbor search and data deduplication. The Hamming distance is a metric used to measure the difference between two binary strings, indicating the number of bits at different positions in the two strings. It is often used in error detection, information encoding and encryption technology. The key pool refers to a set of keys pre-generated and stored in an encryption system for use in subsequent encryption or decryption operations. The key pool can help improve encryption efficiency and the security of key management.

[0132] Construct an updated feature extraction network, using a convolutional neural network structure that includes multiple convolutional layers, pooling layers, and fully connected layers. Input the original user data, such as the user's browsing history, purchase history, social interaction, etc., extract local features through multiple layers of convolution and pooling operations, and then integrate the local features into a global feature vector through a fully connected layer. Use a public key encryption algorithm to encrypt the extracted feature vector to obtain an encrypted feature vector. For example, use the RSA algorithm to encrypt the feature vector, where the public key is public and the private key is held by the data owner.

[0133] The encrypted feature vector is input into the dual autoencoder structure for feature dimensionality reduction. The dual autoencoder is composed of two cascaded autoencoders. The first autoencoder maps the encrypted feature vector to a low-dimensional space to obtain the first compressed feature. The first compressed feature is then input into the decoder of the first autoencoder to reconstruct the initial compressed feature representation. The initial compressed feature representation is then input into the encoder of the second autoencoder to further reduce the feature dimension and obtain the second compressed feature. Finally, the second compressed feature is input into the decoder of the second autoencoder to reconstruct the intermediate feature representation. For example, assuming that the encrypted feature vector dimension is 1024, the first autoencoder compresses it to 512 dimensions, and the second autoencoder further compresses it to 256 dimensions to obtain the intermediate feature representation.

[0134] The intermediate feature representation is input into the sparse autoencoder layer for key feature extraction. By adding the L1 regularization term to the loss function of the sparse autoencoder layer, the activation values ​​of the hidden layer nodes are limited to make them as sparse as possible. The loss function is optimized by the back propagation algorithm to obtain the sparse feature representation. For example, assuming that the intermediate feature representation dimension is 256, the sparse autoencoder layer compresses it to 128 dimensions to obtain a sparse feature representation.

[0135] Use local sensitive hashing function to map sparse feature representation to fixed-length binary feature fingerprint. For example, use SimHash algorithm to convert 128-dimensional sparse feature representation to 64-bit binary feature fingerprint. For any two users' feature fingerprints, count the number of different binary bits corresponding to each feature fingerprint as the Hamming distance to obtain the user similarity matrix. For example, if the feature fingerprints of two users are 10 bits different, the Hamming distance is 10.

[0136] Perform hierarchical clustering based on the user similarity matrix to obtain the initial feature clusters. Use a bottom-up hierarchical clustering algorithm to iteratively merge the most similar user pairs until the preset number of clusters is reached. For example, set the number of clusters to 10, and eventually divide all users into 10 different clusters.

[0137] Construct a multi-layer heterogeneous feature interaction graph. Each user is represented as a node in the graph, and the interaction relationship between users is represented as an edge. User features are divided into multiple dimensions, such as age, gender, hobbies, etc., and a layer of feature interaction subgraph is constructed for each dimension. In each layer of feature interaction subgraph, the weight of the edge is determined based on the feature similarity of the user in the current dimension. For example, if the similarity between two users in the age dimension is high, the weight of the edge connecting the two users is also high.

[0138] The multi-layer heterogeneous feature interaction graph is input into the graph attention network to calculate the attention coefficient between the central node and the neighboring nodes. The neighboring node features are weighted summed based on the attention coefficient, and the node representation is iteratively updated through the multi-layer graph attention network. The user node similarity matrix is ​​calculated based on the updated node representation.

[0139] For each private environment, the user node similarity matrix is ​​partially decrypted according to the pre-assigned sub-key components. For example, the user node similarity matrix is ​​divided into three parts, which are decrypted by three private environments respectively. Through a secure multi-party computing protocol, such as secret sharing, the partial decryption results of each private environment are aggregated to obtain the complete user similarity information.

[0140] Build a hierarchical data access strategy based on the role-based access control model. Divide users into different permission levels, such as ordinary users, advanced users, administrators, etc., and assign corresponding attribute sets to each permission level. For example, the attribute set of ordinary users can be {user}, and the attribute set of advanced users can be {user, VIP}. Generate decryption components that match the attribute set based on the attribute encryption algorithm, and assign corresponding decryption components to users of different permission levels.

[0141] A hierarchical dynamic key negotiation mechanism is used to build a data transmission channel. Independent key pools are maintained at the network layer, transport layer, and application layer. Based on key exchange protocols, such as Diffie-Hellman key exchange, the transmission key is periodically updated between each layer. The complete user similarity information and visitor portrait raw data are securely transmitted between different private domain environments through the data transmission channel.

[0142] In this embodiment, by encrypting feature vectors, multi-party secure computing and hierarchical data access strategies, user privacy is effectively protected and sensitive information leakage is prevented. Dual autoencoders, sparse autoencoder layers and graph attention networks are used to extract user key features, build a multi-layer heterogeneous feature interaction graph, and improve the accuracy and effectiveness of user grouping. A hierarchical dynamic key negotiation mechanism is used to build a secure data transmission channel to ensure that data is securely shared between different private domain environments and prevent data tampering or theft.

[0143] In an optional embodiment,

[0144] The partial decryption results of each private domain environment are aggregated through the secure multi-party computing protocol, and a hierarchical data access strategy is built based on the role-based access control model to divide users into different permission levels, including:

[0145] Using secret sharing technology to divide the partial decryption result of the private domain environment, dividing the partial decryption result into multiple secret shares, and sending the multiple secret shares to other private domain environments and trusted third-party coordinators except the current private domain environment respectively;

[0146] The trusted third-party coordinator receives the multiple secret shares sent by the private domain environment, and combines the secret shares through a secret share reconstruction algorithm to obtain complete user similarity information;

[0147] Building a role-based access control model based on the user's identity information, behavior history, and security level, the role-based access control model including multiple roles and multiple permission sets, each of the roles corresponding to one permission set;

[0148] A rule engine is used to dynamically assign roles to the user, and the user is assigned to one or more roles, and the final permission of the user is the union of the permission sets corresponding to the assigned roles;

[0149] During the data access process, the user's role and permissions are checked in real time, and based on predefined access control rules, it is determined whether the user has the permission to access specific data objects. If the role to which the user belongs has the corresponding permission, the user is allowed to access the data; otherwise, the access operation is not allowed.

[0150] The secret share refers to a secret sharing scheme in which a secret is divided into multiple shares, with each participant holding a share. The original secret can only be restored when all or part of the participants unite. The secret share reconstruction algorithm refers to reconstructing the original secret by merging different secret shares using a mathematical method (such as Lagrange interpolation), which is usually used in applications based on threshold secret sharing. The rule engine refers to a computer system or program that infers and processes input data through a set of preset rules to generate output, and is widely used in automated decision-making, business processes and expert systems.

[0151] Preprocess user data. In each private environment, clean and transform the collected user data, and extract feature vectors for calculating user similarity. For example, user behavior data, including browsing history, purchase history, click behavior, etc., can be converted into numerical vectors. Assume that private environment A has the data of user User1, and its feature vector is [0.2, 0.5, 0.8]; private environment B has the data of user User2, and its feature vector is [0.3, 0.6, 0.7].

[0152] In each private environment, the user's feature vector is partially decrypted. The purpose of partial decryption is to protect the original data while extracting information for calculating similarity. For example, the feature vector can be reduced in dimension or noise can be added. Suppose private environment A partially decrypts the feature vector of User1 and obtains [0.25, 0.55]; private environment B partially decrypts the feature vector of User2 and obtains [0.35, 0.65].

[0153] Use secret sharing technology to split the partial decryption results. Split the partial decryption results of each private environment into multiple secret shares, and distribute these shares to other private environments and trusted third-party coordinators. For example, split [0.25, 0.55] of private environment A into two shares: [0.1, 0.3] and [0.15, 0.25], and send them to private environment B and trusted third-party coordinators respectively. [0.35, 0.65] of private environment B is also split similarly, and the shares are distributed to private environment A and trusted third-party coordinators.

[0154] The trusted third-party coordinator receives secret shares from each private environment and combines them using the secret share reconstruction algorithm to obtain complete user similarity information. For example, the coordinator combines the shares from private environments A and B to reconstruct the partially decrypted feature vectors of User1 and User2, and then calculates the similarity between User1 and User2.

[0155] Build a role-based access control model based on the user's identity information, behavior history, and security level. The model contains multiple roles and multiple permission sets, and each role corresponds to a permission set. For example, define the "ordinary user" role, whose permission set includes "view product information" and "submit order"; define the "VIP user" role, whose permission set includes "view product information", "submit order", and "enjoy exclusive discounts".

[0156] Use the rule engine to dynamically assign roles to users. According to pre-set rules, users are assigned to one or more roles. The final permissions of the user are the union of the permission sets corresponding to the assigned roles. For example, based on the user's consumption amount, users with a consumption amount greater than 10,000 are assigned to the "VIP user" role, and other users are assigned to the "ordinary user" role.

[0157] During data access, the user's role and permissions are checked in real time. Based on the predefined access control rules, it is determined whether the user has the permission to access a specific data object. If the role to which the user belongs has the corresponding permission, the user is allowed to access the data, otherwise the access is denied. For example, when a user tries to access the "Enjoy Exclusive Discount" function, the system checks the user's role. If it is a "VIP user", access is allowed; if it is a "normal user", access is denied.

[0158] In this embodiment, secret sharing and partial decryption technology are used to protect the user's original data and prevent data leakage. Even if part of the private domain environment or the trusted third-party coordinator is compromised, the complete user data cannot be obtained. Through the secure multi-party computing protocol, user similarity calculation is achieved without leaking the original data, which improves the efficiency of data sharing and avoids the complex process of data transmission and integration. Through the role-based access control model, refined management of user access rights is achieved. Access rights can be flexibly configured according to different user attributes and security levels to ensure data security and availability.

[0159] Figure 2 Schematic diagram of the structure of a multi-private domain visitor portrait sharing and privacy protection routing system based on federated learning in an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0160] The first unit is used to obtain the original data of visitor portraits in multiple independent private domain environments, calculate the information entropy corresponding to each category of data in the original data of the visitor portraits based on JS divergence, obtain a quantitative index of feature sensitivity, calculate the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtain the differential privacy budget allocation result and construct a mixed noise generator, construct a mixed noise sequence and add it to the original data of the visitor portraits, generate a privacy-preserving feature vector, generate a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypt the privacy-preserving feature vector to obtain an encrypted feature vector, perform threshold cryptographic segmentation on the fully homomorphic encryption key through a Viterbi algorithm, obtain multiple sub-key components, and send the sub-key components to the independent private domain environment through a distributed consistency protocol;

[0161] The second unit is used to build a hybrid network structure based on a multi-level recursive neural network and a graph neural network, add an adaptive attention mechanism and a residual connection mechanism to each layer of the hybrid network structure to obtain a feature extraction network, add the encrypted feature vector to the feature extraction network to obtain an initial feature representation, determine a data quality assessment index based on the initial feature representation, calculate the historical contribution of each independent private domain environment through a time series weighted decay algorithm, weightedly fuse the data quality assessment index with the historical contribution to obtain a model aggregation weight, build a two-layer asynchronous federated optimization framework, update and detect anomalies on the feature extraction network based on the model aggregation weight, determine effective local model parameters and obtain an updated feature extraction network;

[0162] The third unit is used to add the updated feature extraction network and the encrypted feature vector to the dual autoencoder structure to perform feature dimensionality reduction to obtain an intermediate feature representation, extract key features through a sparse autoencoder layer and generate a feature fingerprint corresponding to each user according to a preset feature fingerprint generation function, calculate the similarity between different feature fingerprints and cluster them to obtain an initial feature cluster, execute an information loss minimization algorithm to group users, calculate a user node similarity matrix based on a multi-layer heterogeneous feature interaction graph and a graph attention network, partially decrypt the user node similarity matrix based on the sub-key component, aggregate the partial decryption results corresponding to each independent private domain environment in combination with a multi-party secure computing protocol, obtain complete user similarity information, and assign a decryption component to each user according to a pre-built hierarchical data access strategy, and securely share the data transmission channel built based on a hierarchical dynamic key negotiation mechanism.

[0163] According to a third aspect of the embodiments of the present invention,

[0164] An electronic device is provided, comprising:

[0165] processor;

[0166] a memory for storing processor-executable instructions;

[0167] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0168] According to a fourth aspect of the embodiments of the present invention,

[0169] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0170] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-private domain visitor portrait sharing and privacy protection routing method based on federated learning, characterized in that: include: Obtain original visitor portrait data in multiple independent private domain environments, calculate the information entropy corresponding to each category of data in the original visitor portrait data based on JS divergence, obtain a quantitative indicator of feature sensitivity, calculate the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtain the differential privacy budget allocation result and construct a mixed noise generator, construct a mixed noise sequence and add it to the original visitor portrait data to generate a privacy-preserving feature vector, generate a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypt the privacy-preserving feature vector to obtain an encrypted feature vector, perform threshold cryptographic segmentation on the fully homomorphic encryption key through the Viterbi algorithm, obtain multiple sub-key components, and send the sub-key components to the independent private domain environment through a distributed consistency protocol; A hybrid network structure is constructed based on a multi-level recursive neural network and a graph neural network, an adaptive attention mechanism and a residual connection mechanism are added to each layer of the hybrid network structure to obtain a feature extraction network, the encrypted feature vector is added to the feature extraction network to obtain an initial feature representation, a data quality assessment index is determined based on the initial feature representation, the historical contribution of each independent private domain environment is calculated by a time series weighted decay algorithm, the data quality assessment index is weightedly fused with the historical contribution to obtain a model aggregation weight, a two-layer asynchronous federated optimization framework is constructed, and the feature extraction network is updated and anomaly detected based on the model aggregation weight, and effective local model parameters are determined to obtain an updated feature extraction network; The updated feature extraction network and encrypted feature vector are added to the dual autoencoder structure to perform feature dimensionality reduction to obtain intermediate feature representation, key features are extracted through the sparse autoencoder layer, and the feature fingerprint corresponding to each user is generated according to the preset feature fingerprint generation function, the similarity between different feature fingerprints is calculated and clustered to obtain the initial feature cluster, and the information loss minimization algorithm is executed to group users, and the user node similarity matrix is ​​calculated based on the multi-layer heterogeneous feature interaction graph and the graph attention network. The user node similarity matrix is ​​partially decrypted based on the sub-key component, and the partial decryption results corresponding to each independent private domain environment are aggregated in combination with the multi-party secure computing protocol to obtain complete user similarity information and assign decryption components to each user according to the pre-built hierarchical data access strategy, which is combined with the data transmission channel built based on the hierarchical dynamic key negotiation mechanism for secure sharing.

2. The method according to claim 1, characterized in that Obtaining original visitor portrait data in multiple independent private domain environments, calculating the information entropy corresponding to each category of data in the original visitor portrait data based on JS divergence, obtaining a quantitative indicator of feature sensitivity, calculating the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtaining the differential privacy budget allocation result and constructing a mixed noise generator, constructing a mixed noise sequence and adding it to the original visitor portrait data, generating a privacy-preserving feature vector, generating a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypting the privacy-preserving feature vector to obtain an encrypted feature vector, performing threshold cryptographic segmentation on the fully homomorphic encryption key through a Viterbi algorithm, obtaining multiple sub-key components, and sending the sub-key components to the independent private domain environment through a distributed consistency protocol, including: Obtaining original visitor portrait data in multiple independent private domain environments, wherein the original visitor portrait data includes user ID, age, gender, region, browsing history, and purchase history; Calculate the probability distribution of each data category in the original data of the visitor portrait and the average distribution of the overall data, substitute the probability distribution of the data category and the average distribution into the Kuller-Leibler divergence formula to calculate a first divergence value, substitute the probability distribution of the overall data and the average distribution into the Kuller-Leibler divergence formula to calculate a second divergence value, perform weighted averaging of the first divergence value and the second divergence value to obtain the Jensen-Shannon divergence, and normalize the Jensen-Shannon divergence to obtain a feature sensitivity quantification index; Construct a differential privacy budget allocation model based on dynamic programming, set the total privacy budget as the target constraint, set the feature sensitivity quantification indicator as the optimization weight, construct a state transfer equation, where the state variable of the state transfer equation is the allocated privacy budget, and the decision variable is the privacy budget to be allocated. Iterate and calculate the optimal substructure based on the state transfer equation to obtain the privacy budget allocation value for each data category; Constructing a mixed noise generator, the mixed noise generator including a Laplace noise module and a Gaussian noise module, calculating the numerical sensitivity of each of the data categories, substituting the numerical sensitivity and the corresponding privacy budget allocation value into the Laplace distribution function and the Gaussian distribution function respectively, calculating the noise intensity corresponding to each of the data categories based on the Laplace distribution function and the Gaussian distribution function, adding random noise to each of the data categories through the mixed noise generator according to the noise intensity, obtaining a mixed noise sequence, and adding the mixed noise sequence to the original visitor portrait data to obtain a privacy protection feature vector; Constructing security parameters of an ideal lattice cryptographic system, generating two large integers as public keys based on the security parameters, randomly selecting a polynomial as a private key, combining the public key and the private key to form the fully homomorphic encryption key pair, selecting a random noise polynomial, and using the public key in the fully homomorphic encryption key pair to perform encryption operation on the privacy protection feature vector to obtain an encrypted feature vector; The private key in the fully homomorphic encryption key pair is coefficient decomposed to obtain a coefficient vector, a polynomial interpolation operation is performed on the coefficient vector to obtain multiple coordinate points, the coordinate points are secretly divided by a Viterbi algorithm to obtain multiple sub-key components, and the multiple sub-key components are respectively distributed to each of the independent private domain environments through a distributed consistency protocol.

3. The method according to claim 2, characterized in that The coordinate points are secretly divided by the Viterbi algorithm to obtain multiple sub-key components including: Randomly select two prime numbers greater than 100 and multiply them to obtain a modulus, calculate the lowest common multiple of the prime numbers minus one to obtain a private key parameter, select a generator with a specified order, the order of the generator is the square of the modulus, and combine the modulus and the generator to obtain a public key; Select a random number for each coordinate point, perform modular exponentiation operation on the horizontal coordinate and the vertical coordinate of the coordinate point, the generator and the random number, respectively, to obtain an encrypted coordinate point, the encrypted coordinate point including an encrypted horizontal coordinate and an encrypted vertical coordinate, and construct all the encrypted coordinate points into an encrypted coordinate point matrix; Constructing a state transfer equation, wherein the input variable of the state transfer equation is the encrypted coordinate point, and the output variable of the state transfer equation is the encrypted path score, performing dynamic programming calculation on the ciphertext domain, recursively calculating the optimal path score of each state node based on the state transfer equation, and obtaining a global optimal encrypted path by reverse backtracking, wherein the global optimal encrypted path is composed of multiple encrypted coordinate points connected in a transfer order; Divide the encrypted coordinate point matrix based on the global optimal encryption path, divide the encrypted coordinate point matrix into multiple encryption subsets, each encryption subset contains the encrypted coordinate points of the corresponding state in the global optimal encryption path, construct a random polynomial, the constant term of the random polynomial is the sum of the encrypted coordinate points in the encryption subset, and the other coefficients of the random polynomial are encrypted random numbers, calculate the values ​​of the random polynomial at different input values, obtain multiple encryption coefficients, and distribute the encryption coefficients as subkey components to different participants; Receive the encrypted subkey components returned by the participant, aggregate the encrypted subkey components, decrypt the aggregated results according to the private key parameters to obtain the plaintext subkey components, substitute the plaintext subkey components into the secret sharing recovery function, and restore the complete private key.

4. The method according to claim 1, characterized in that: A hybrid network structure is constructed based on a multi-level recursive neural network and a graph neural network, an adaptive attention mechanism and a residual connection mechanism are added to each layer of the hybrid network structure to obtain a feature extraction network, the encrypted feature vector is added to the feature extraction network to obtain an initial feature representation, a data quality evaluation index is determined based on the initial feature representation, the historical contribution of each independent private domain environment is calculated by a time series weighted decay algorithm, the data quality evaluation index and the historical contribution are weightedly fused to obtain a model aggregation weight, a two-layer asynchronous federated optimization framework is constructed, and the feature extraction network is updated and anomaly detected based on the model aggregation weight, and effective local model parameters are determined to obtain an updated feature extraction network, including: Based on multi-level recursive neural networks and graph neural networks, multiple long short-term memory network layers and multiple graph convolutional network layers are alternately stacked to construct a hybrid network structure; Calculating the dot product attention weight matrix of each layer of the hybrid network structure, multiplying the attention weight matrix with the input feature to obtain a weighted feature, adding the weighted feature to the input feature to generate a residual enhancement feature, and repeatedly calculating the weighted feature and the residual enhancement feature in each layer of the hybrid network structure to obtain a feature extraction network; Receiving an encrypted feature vector, inputting the encrypted feature vector into the hybrid network structure, outputting an initial feature representation, extracting a missing value ratio, an outlier ratio, and data redundancy in the initial feature representation, generating a quality assessment score, and performing a weighted average calculation on the quality assessment score to obtain a comprehensive data quality assessment index; Obtain the historical gradient contribution sequence of each private domain environment, generate an attenuation weight sequence based on the time step, perform a weighted sum operation on the historical gradient contribution sequence and the attenuation weight sequence according to the time series weighted attenuation algorithm, calculate the historical contribution corresponding to each private domain environment, perform a weighted fusion operation on the comprehensive data quality assessment index and the historical contribution, and generate a model aggregation weight for each private domain environment; Construct a two-layer asynchronous federated optimization framework, wherein the two-layer asynchronous federated optimization framework includes a global optimization layer and a local optimization layer. The global optimization layer collects uploaded local model parameters from each private domain environment, and performs weighted averaging according to the corresponding model aggregation weights to obtain global model parameters. The global model parameters are transmitted from the central server to the local device of each private domain environment through a secure communication protocol. In the local optimization layer, each private domain environment updates the global model parameters according to local data through a gradient descent algorithm to obtain updated local model parameters that adapt to the local data distribution. The gradient norm difference value between the updated local model parameters and the global model parameters is calculated, the local model parameters whose gradient norm difference value exceeds a preset threshold are marked as abnormal parameters and filtered, and the normal parameters are retained for parameter update of the feature extraction network to obtain an updated feature extraction network.

5. The method according to claim 4, characterized in that The calculation formula corresponding to the historical contribution is as follows: ； in, c j Represents private domain environment j The historical contribution of T represents the total number of time steps, t represents the time step index, λ 0 represents the initial attenuation coefficient, α j ( t ) indicates the private domain environment j At time step t data quality, β represents the contribution rate adjustment coefficient, δ j ( t ) indicates the private domain environment j At time step t The gradient change rate, γ represents the gradient influence coefficient, g j ( t ) indicates the private domain environment j At time step t The gradient contribution of .

6. The method according to claim 1, characterized in that The updated feature extraction network and encrypted feature vector are added to the dual autoencoder structure to perform feature dimension reduction to obtain intermediate feature representation, key features are extracted through the sparse autoencoder layer and feature fingerprints corresponding to each user are generated according to a preset feature fingerprint generation function, the similarity between different feature fingerprints is calculated and clustered to obtain the initial feature cluster, and the information loss minimization algorithm is executed to group users, and the user node similarity matrix is ​​calculated based on the multi-layer heterogeneous feature interaction graph and the graph attention network, and the user node similarity matrix is ​​partially decrypted based on the sub-key component, and the partial decryption results corresponding to each independent private domain environment are aggregated in combination with the multi-party secure computing protocol to obtain complete user similarity information and assign decryption components to each user according to the pre-built hierarchical data access strategy, and the data transmission channel built based on the hierarchical dynamic key negotiation mechanism is combined for secure sharing, including: Inputting the updated feature extraction network and the encrypted feature vector into a dual autoencoder structure, the dual autoencoder structure comprising a first autoencoder and a second autoencoder, inputting the encrypted feature vector into the encoder of the first autoencoder, mapping the encrypted feature vector to a low-dimensional space through a multi-layer feedforward neural network to obtain a first compressed feature, inputting the first compressed feature into the decoder of the first autoencoder to reconstruct a preliminary compressed feature representation, inputting the preliminary compressed feature representation into the encoder of the second autoencoder to reduce the feature dimension to obtain a second compressed feature, and inputting the second compressed feature into the decoder of the second autoencoder to reconstruct an intermediate feature representation; The intermediate feature representation is input into a sparse autoencoder layer, the activation sparsity of the hidden layer nodes is modified by adding a norm regularization term to the loss function of the sparse autoencoder layer, the loss function is optimized by a back propagation algorithm to obtain a sparse feature representation, a local sensitive hash function is used as a feature fingerprint generation function, the sparse feature representation is mapped into a binary feature fingerprint of a fixed length, for the feature fingerprints of any two users, the number of different binary bits corresponding to each feature fingerprint is counted as the Hamming distance to obtain a user similarity matrix, and the most similar user pairs are iteratively merged based on the user similarity matrix through a bottom-up hierarchical clustering algorithm until a preset number of clusters is reached to obtain an initial feature cluster; Construct a multi-layer heterogeneous feature interaction graph, where each user is represented as a node in the graph, and the interaction relationship between users is represented as an edge. The user features are divided into multiple dimensions, and a layer of feature interaction subgraph is constructed for each dimension. In the feature interaction subgraph corresponding to each layer, the weight of the edge is determined based on the feature similarity of the user in the current dimension. The multi-layer heterogeneous feature interaction graph is input into the graph attention network, and the attention coefficient between the central node and the neighboring nodes is calculated. The neighboring node features are weighted summed based on the attention coefficient. The node representation is iteratively updated through the multi-layer graph attention network, and the user node similarity matrix is ​​calculated based on the updated node representation. For each private domain environment, partially decrypt the user node similarity matrix according to the pre-assigned sub-key components, aggregate the partial decryption results of each private domain environment through a secure multi-party computing protocol, build a hierarchical data access strategy based on a role-based access control model, divide users into different permission levels, assign a corresponding attribute set to each permission level, generate a decryption component matching the attribute set based on an attribute encryption algorithm, and assign the corresponding decryption component to users at different permission levels; A hierarchical dynamic key negotiation mechanism is adopted to build a data transmission channel, and independent key pools are maintained at the network, transmission and application levels respectively. The transmission key is periodically updated between the levels based on the key exchange protocol, and the complete user similarity information and the original visitor portrait data are securely transmitted between different private domain environments through the data transmission channel.

7. The method according to claim 6, characterized in that The partial decryption results of each private domain environment are aggregated through the secure multi-party computing protocol, and a hierarchical data access strategy is built based on the role-based access control model to divide users into different permission levels, including: Using secret sharing technology to divide the partial decryption result of the private domain environment, dividing the partial decryption result into multiple secret shares, and sending the multiple secret shares to other private domain environments and trusted third-party coordinators except the current private domain environment respectively; The trusted third-party coordinator receives the multiple secret shares sent by the private domain environment, and combines the secret shares through a secret share reconstruction algorithm to obtain complete user similarity information; Building a role-based access control model based on the user's identity information, behavior history, and security level, the role-based access control model including multiple roles and multiple permission sets, each of the roles corresponding to one permission set; A rule engine is used to dynamically assign roles to the user, and the user is assigned to one or more roles, and the final permission of the user is the union of the permission sets corresponding to the assigned roles; During the data access process, the user's role and permissions are checked in real time, and based on predefined access control rules, it is determined whether the user has the permission to access specific data objects. If the role to which the user belongs has the corresponding permission, the user is allowed to access the data; otherwise, the access operation is not allowed.

8. A multi-private domain visitor portrait sharing and privacy protection routing system based on federated learning, used to implement any of the methods in claims 1-7, characterized in that: include: The first unit is used to obtain the original data of visitor portraits in multiple independent private domain environments, calculate the information entropy corresponding to each category of data in the original data of the visitor portraits based on JS divergence, obtain a quantitative index of feature sensitivity, calculate the optimal allocation plan of the differential privacy budget through a dynamic programming algorithm, obtain the differential privacy budget allocation result and construct a mixed noise generator, construct a mixed noise sequence and add it to the original data of the visitor portraits, generate a privacy-preserving feature vector, generate a fully homomorphic encryption key pair through a public key encryption algorithm based on an ideal lattice and encrypt the privacy-preserving feature vector to obtain an encrypted feature vector, perform threshold cryptographic segmentation on the fully homomorphic encryption key through a Viterbi algorithm, obtain multiple sub-key components, and send the sub-key components to the independent private domain environment through a distributed consistency protocol; The second unit is used to build a hybrid network structure based on a multi-level recursive neural network and a graph neural network, add an adaptive attention mechanism and a residual connection mechanism to each layer of the hybrid network structure to obtain a feature extraction network, add the encrypted feature vector to the feature extraction network to obtain an initial feature representation, determine a data quality assessment index based on the initial feature representation, calculate the historical contribution of each independent private domain environment through a time series weighted decay algorithm, weightedly fuse the data quality assessment index with the historical contribution to obtain a model aggregation weight, build a two-layer asynchronous federated optimization framework, update and detect anomalies on the feature extraction network based on the model aggregation weight, determine effective local model parameters and obtain an updated feature extraction network; The third unit is used to add the updated feature extraction network and the encrypted feature vector to the dual autoencoder structure to perform feature dimensionality reduction to obtain an intermediate feature representation, extract key features through a sparse autoencoder layer and generate a feature fingerprint corresponding to each user according to a preset feature fingerprint generation function, calculate the similarity between different feature fingerprints and cluster them to obtain an initial feature cluster, execute an information loss minimization algorithm to group users, calculate a user node similarity matrix based on a multi-layer heterogeneous feature interaction graph and a graph attention network, partially decrypt the user node similarity matrix based on the sub-key component, aggregate the partial decryption results corresponding to each independent private domain environment in combination with a multi-party secure computing protocol, obtain complete user similarity information, and assign a decryption component to each user according to a pre-built hierarchical data access strategy, and securely share the data transmission channel built based on a hierarchical dynamic key negotiation mechanism.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images