Privacy protection method and system based on big data security and privacy computing

Abstract

The present invention provides a privacy protection method and system based on big data security and privacy computing, which relates to the field of data processing technology, including: obtaining original data feature extraction, obtaining data features and selecting a data desensitization algorithm to perform data desensitization, performing encryption processing through a hybrid fully homomorphic encryption algorithm, generating homomorphically encrypted ciphertext data and sending it to a privacy protection computing platform; constructing a privacy protection data preprocessing protocol, performing data filtering and data dimension compression, executing multi-party secure computing, generating ciphertext computing results and returning them to a data provider for verification, and if the verification passes, decrypting them through a decryption algorithm, filtering to obtain pure plaintext computing and analysis results; constructing a reverse data desensitization model for restoration, obtaining analysis and calculation results, constructing and solving a dual-objective optimization model, obtaining an optimal balance point and generating an optimal decision result, performing desensitization, uploading to a privacy protection blockchain network, and performing distributed trusted storage and sharing.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a privacy protection method and system based on big data security and privacy computing. Background Art

[0002] In today's information age, with the rapid development and widespread application of big data technology, data security and privacy protection have become a key issue. The core of big data technology lies in data collection, storage, processing and analysis, but in these processes, the risk of data privacy leakage has also increased significantly. In order to fully utilize the value of big data and protect user privacy at the same time, privacy protection technology based on big data security and privacy computing has emerged;

[0003] Although traditional privacy protection methods can protect data privacy to a certain extent, they still have some limitations, such as reduced data availability, difficulty in supporting complex analytical calculations, and the inability of encrypted data to directly participate in calculations. It needs to be decrypted before calculation, which poses the risk of data leakage and performance loss. In addition, most existing privacy-preserving data mining methods are aimed at structured data, and lack support for privacy-preserving mining of unstructured data and heterogeneous data.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a privacy protection method and system based on big data security and privacy computing, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a privacy protection method based on big data security and privacy computing, comprising:

[0007] Obtaining the original data to be processed, extracting features through a data feature extraction model based on deep learning, obtaining data features and selecting a data desensitization algorithm according to the data features, desensitizing the data through the selected data desensitization algorithm, encrypting the desensitized privacy data through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, generating homomorphically encrypted ciphertext data and sending it to a privacy protection computing platform;

[0008] After receiving the ciphertext data, the privacy-preserving computing platform constructs a privacy-preserving data preprocessing protocol based on locally sensitive hashing and secret sharing technology, performs data filtering and data dimension compression in the ciphertext domain, and performs multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on verifiable random functions, generates ciphertext computing results and returns them to the data provider, and the data provider verifies the data through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts the data through a decryption algorithm to generate a plaintext computing analysis result, and filters the data through adaptive filtering theory to obtain a pure plaintext computing analysis result;

[0009] Based on the pure plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network. The pure plaintext calculation and analysis results are restored through reverse learning and adversarial generation to obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate the optimal decision result. The optimal decision result is desensitized by the data desensitization algorithm, and the desensitized optimal decision result is uploaded to the privacy protection blockchain network for distributed trusted storage and sharing.

[0010] In an optional embodiment,

[0011] The raw data to be processed is obtained, and features are extracted through a data feature extraction model based on deep learning to obtain data features and select a data desensitization algorithm according to the data features, and data desensitization is performed through the selected data desensitization algorithm. For the desensitized privacy data, the desensitized privacy data is encrypted through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, and homomorphically encrypted ciphertext data is generated and sent to the privacy protection computing platform, including:

[0012] The raw data to be processed is obtained and added to a data feature extraction model based on deep learning. The data feature extraction model determines the data type through a discriminator, and divides the raw data into image data, sequence data, and graph data. For the image data, feature extraction is performed through a convolutional layer and a pooling layer of a pre-set convolutional neural network. For the sequence data, feature extraction is performed on hidden state updates through a recurrent neural network. For the graph data, feature extraction is performed through node updates and message passing in a graph neural network. The extracted features are feature fused to obtain the data features.

[0013] Construct a data desensitization algorithm library containing multiple data desensitization algorithms, quantitatively evaluate the data utility loss of each algorithm in the data desensitization algorithm library in combination with a pre-set utility measurement indicator, determine the desensitization utility corresponding to each data desensitization algorithm, and based on the desensitization utility and privacy protection effect, construct and train a desensitization decision tree model with data privacy attribute characteristics and privacy protection requirements as input and the data desensitization algorithm as output, extract the privacy attribute characteristics of the data features and add them to the desensitization decision tree model, obtain the data desensitization algorithm with the highest matching degree with the data features, and use the algorithm for data desensitization;

[0014] The desensitized data features are divided into real data and integer data, the real data are encoded into integers based on the lattice cryptography system and the integers are embedded into the lattice space through homomorphic mapping to generate lattice ciphertext, for the integer data, a fully homomorphic encryption scheme is selected through a fully homomorphic encryption algorithm, a public-private key pair is generated according to the parameters of the fully homomorphic encryption scheme and homomorphic encryption is performed to obtain a fully homomorphic encrypted ciphertext, the lattice ciphertext and the fully homomorphic encrypted ciphertext are uniformly encapsulated to generate a mixed fully homomorphic ciphertext data packet and sent to the privacy protection computing platform.

[0015] In an optional embodiment,

[0016] Combined with the preset utility measurement indicators, the data utility loss of each algorithm in the data desensitization algorithm library is quantitatively evaluated as shown in the following formula:

[0017]

[0018] Among them, LOSS represents data utility loss, α represents penalty factor, wi represents the weight of the i-th performance indicator, and Perf i () represents the i-th performance evaluation index, M Y represents the machine learning model trained on the desensitized data D', M X represents the machine learning model trained on the original data D, β represents the balance loss weight, Bal(D′) represents the balance of the desensitized data D′, and Bal(D) represents the balance of the original data D.

[0019] In an optional embodiment,

[0020] After receiving the ciphertext data, the privacy-preserving computing platform constructs a privacy-preserving data preprocessing protocol based on locally sensitive hashing and secret sharing technology, performs data filtering and data dimension compression in the ciphertext domain, and performs multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on verifiable random functions. The ciphertext computing result is generated and returned to the data provider, and the data provider verifies it through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts it through a decryption algorithm to generate a plaintext computing analysis result, and filters it through adaptive filtering theory to obtain a pure plaintext computing analysis result, including:

[0021] After receiving the ciphertext data, the privacy protection computing platform constructs a data preprocessing protocol, wherein the data preprocessing protocol includes a data filtering step and a data dimension compression step;

[0022] In the data filtering step, according to the format of the ciphertext data, multiple groups of hash function parameters are selected based on local sensitive hashing and multiple hash functions are generated, the values ​​of the hash functions are used as the hash signatures of each data point in the ciphertext data, the signature similarity between different hash signatures is calculated and compared with a preset similarity threshold, if the signature similarity is greater than the similarity threshold, the data point corresponding to the current hash signature is filtered, and the comparison is repeated until all data points are processed to obtain a set of retained data points;

[0023] In the data dimension compression step, for each element in the reserved data point set, a polynomial is constructed, the value of the polynomial is decomposed according to the number of participants, and a plurality of low-dimensional secret shares are generated and sent to the participants;

[0024] After receiving the low-dimensional secret share, the participant generates a key pair based on a verifiable random function and broadcasts the public key to other participants. For the current computing task, each participant sends its ciphertext input and public key to the privacy protection computing platform. After collecting all ciphertext inputs and public keys, the privacy protection computing platform randomly selects a participant as the computing initiator. The computing initiator generates a random seed and calculates the output and proof of the verifiable random function. The random seed, the output of the verifiable random function and the proof are sent to the privacy protection computing platform and the correctness of the output is verified using the public key. If the verification is successful, a random value is selected as the seed of the pseudo-random number generator to generate a random mask. The random mask is homomorphically operated with the ciphertext input to obtain a mask calculation result. Based on the mask calculation result, each participant decrypts using a private key and sends the decrypted masked plaintext result to the computing initiator. The ciphertext calculation result is obtained by removing the random mask and returned to the data provider.

[0025] After the data provider receives the ciphertext calculation result, the privacy-preserving computing platform generates a public reference string based on a non-interactive zero-knowledge proof scheme and sends it to each participant. At the same time, the ciphertext calculation result is used as a statement, and the random mask is used as evidence. A zero-knowledge proof is generated by a generation proof algorithm and sent to the data provider. The data provider verifies according to the verification algorithm. If the verification passes, the data provider uses a decryption method based on homomorphic encryption to decrypt the ciphertext calculation result to obtain the plaintext calculation analysis result, initialize the pre-set filter weights and calculate the filtering error, adaptively update the filter weights according to the filtering error, and filter the plaintext calculation analysis result to obtain the pure plaintext calculation analysis result.

[0026] In an optional embodiment,

[0027] The weight of the filter is adaptively updated according to the filtering error as shown in the following formula:

[0028]

[0029] Among them, w(n) represents the filter weight vector at time step n, w(n-1) represents the filter weight vector at time step n-1, P(n-1) represents the filter prediction error covariance matrix at time step n-1, and m r (n) represents the input vector at time step n, w T (n-1) represents the transpose of the filter weight vector at time step n-1, represents the transpose of the input vector at time step n, and λ represents the forgetting rate.

[0030] In an optional embodiment,

[0031] Based on the pure plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network. The pure plaintext calculation and analysis results are restored through reverse learning and adversarial generation to obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate the optimal decision result. The optimal decision result is desensitized by the data desensitization algorithm, and the desensitized optimal decision result is uploaded to the privacy protection blockchain network for distributed trusted storage and sharing, including:

[0032] Based on the clean plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network, wherein the data desensitization model includes a desensitizer and a restorer, the desensitizer adopts the structure of the generator in the deep generative adversarial network, and generates desensitized data by performing a deconvolution operation on the clean plaintext calculation and analysis results. For the desensitized data, the restorer adopts the structure of the discriminator in the deep generative adversarial network, restores the desensitized data through a convolution operation, outputs the restored data, repeats the iteration and adjusts the hyperparameters of the desensitizer and restorer according to the original data corresponding to the clean plaintext calculation and analysis results until convergence or reaching the preset maximum number of iterations, and uses the restored data output by the last iteration as the analysis and calculation result;

[0033] Based on the analysis and calculation results, the data provider constructs a dual-objective optimization model with privacy protection strength and decision effectiveness as the goals according to a multi-objective evolutionary optimization algorithm, adds the analysis and calculation results to the dual-objective optimization model, randomly generates an initial population based on the analysis and calculation results, performs non-dominated sorting on the individuals in the initial population, obtains the non-dominated level and crowding degree corresponding to each individual, selects multiple individuals from the initial population to form a parent population through a binary tournament selection operator, performs crossover and mutation operations on the parent population to obtain a child population, merges the child population and the initial population to generate a candidate second-generation population and performs non-dominated sorting to obtain the dominance relationship and crowding degree between individuals in the candidate second-generation population, removes the dominated individuals, and for the retained individuals, sorts them in descending order according to the crowding degree of each individual, selects the top 10 individuals to retain, obtains the second-generation population, repeats the iteration until the preset maximum number of iterations is reached, selects the individual ranked first in the last iteration as the optimal balance point, and generates the optimal decision result based on the privacy protection strength and decision effectiveness corresponding to the optimal balance point;

[0034] The optimal decision result is desensitized by the data desensitization algorithm, the desensitized optimal decision result is digitally signed and packaged and uploaded to the privacy protection blockchain network, the block is verified by the consensus mechanism and the verified block is added to the local blockchain to complete distributed trusted storage and sharing.

[0035] In an optional embodiment,

[0036] The crowding degree of each individual is calculated in the non-dominated sorting process as shown in the following formula:

[0037]

[0038] Among them, d grepresents the crowding degree of individual g, K represents the number of objective functions, represents the next objective function value in the non-dominated level where individual g is located on the kth objective function, represents the previous objective function value in the non-dominated level of individual g on the kth objective function, represents the maximum value of the kth objective function in the current non-dominated level, represents the minimum value of the kth objective function in the current non-dominated level, x g represents the decision variable vector corresponding to individual g, x h represents the decision variable vector corresponding to individual h, N represents the population size in the individual, and ‖‖ represents the Euclidean distance.

[0039] A second aspect of an embodiment of the present invention provides a privacy protection system based on big data security and privacy computing, including:

[0040] The first unit is used to obtain the original data to be processed, perform feature extraction through a data feature extraction model based on deep learning, obtain data features and select a data desensitization algorithm according to the data features, perform data desensitization through the selected data desensitization algorithm, encrypt the desensitized privacy data through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, generate homomorphically encrypted ciphertext data and send it to a privacy protection computing platform;

[0041] The second unit is used for the privacy-preserving computing platform to construct a privacy-preserving data preprocessing protocol based on local sensitive hashing and secret sharing technology after receiving the ciphertext data, perform data filtering and data dimension compression in the ciphertext domain, perform multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on a verifiable random function, generate a ciphertext computing result and return it to the data provider, and the data provider verifies it through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts it through a decryption algorithm to generate a plaintext computing analysis result, and filters it through adaptive filtering theory to obtain a pure plaintext computing analysis result;

[0042] The third unit is used to construct a reverse data desensitization model based on the pure plaintext calculation and analysis results in combination with the data desensitization algorithm and the deep generative adversarial network, restore the pure plaintext calculation and analysis results through reverse learning and adversarial generation, and obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate an optimal decision result, desensitizes the optimal decision result through the data desensitization algorithm, and uploads the desensitized optimal decision result to the privacy protection blockchain network for distributed trusted storage and sharing.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the embodiments of the present invention is:

[0049] 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.

[0050] In the present invention, the features of the original data are extracted through a data feature extraction model based on deep learning, and an appropriate data desensitization algorithm is selected according to the features to perform data desensitization. The desensitized privacy data is encrypted through a hybrid fully homomorphic encryption algorithm to generate homomorphically encrypted ciphertext data, thereby realizing privacy protection of the original data and preventing the risk of data leakage during transmission and calculation. Data filtering and data dimension compression are performed in the ciphertext domain, reducing the scale and complexity of calculation. The ciphertext calculation results are returned to the data provider for decryption and verification, ensuring the correctness and credibility of the calculation results. Through reverse learning and adversarial generation, the pure plaintext calculation is protected. The calculation and analysis results are restored to obtain the analysis and calculation results corresponding to the original data, thereby realizing the privacy protection and restoration of the calculation results, ensuring the data provider's control and ownership of the original data, and realizing distributed trusted storage and secure sharing of decision results through the distributed consensus mechanism and encryption technology of blockchain. In summary, this embodiment realizes the full-process privacy protection and secure sharing of original data, calculation process, calculation results and decision results, ensuring the privacy security and data ownership of the data provider, improving the data utilization efficiency and collaboration level among multiple parties, and providing strong technical support for the practical application of privacy protection data mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flowchart of a privacy protection method based on big data security and privacy computing according to an embodiment of the present invention;

[0052] Figure 2 This is a structural diagram of a privacy protection system based on big data security and privacy computing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] 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.

[0054] 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.

[0055] Figure 1 FIG. 1 is a flow chart of a privacy protection method based on big data security and privacy computing according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] S1. Obtain the original data to be processed, extract features through a data feature extraction model based on deep learning, obtain data features and select a data desensitization algorithm according to the data features, desensitize the data through the selected data desensitization algorithm, encrypt the desensitized privacy data through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, generate homomorphically encrypted ciphertext data and send it to the privacy protection computing platform;

[0057] The data desensitization algorithm is designed to protect sensitive data and ensure that the data remains useful to legitimate users while ensuring the privacy of the data during use. The lattice cryptographic system is based on lattice theory and provides an encryption method that is resistant to quantum computing attacks. The fully homomorphic encryption is an encryption method that allows encrypted data to be directly operated on without decryption. The hybrid fully homomorphic encryption algorithm combines the advantages of symmetric encryption and fully homomorphic encryption.

[0058] In an optional embodiment,

[0059] The raw data to be processed is obtained, and features are extracted through a data feature extraction model based on deep learning to obtain data features and select a data desensitization algorithm according to the data features, and data desensitization is performed through the selected data desensitization algorithm. For the desensitized privacy data, the desensitized privacy data is encrypted through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, and homomorphically encrypted ciphertext data is generated and sent to the privacy protection computing platform, including:

[0060] The raw data to be processed is obtained and added to a data feature extraction model based on deep learning. The data feature extraction model determines the data type through a discriminator, and divides the raw data into image data, sequence data, and graph data. For the image data, feature extraction is performed through a convolutional layer and a pooling layer of a pre-set convolutional neural network. For the sequence data, feature extraction is performed on hidden state updates through a recurrent neural network. For the graph data, feature extraction is performed through node updates and message passing in a graph neural network. The extracted features are feature fused to obtain the data features.

[0061] Construct a data desensitization algorithm library containing multiple data desensitization algorithms, quantitatively evaluate the data utility loss of each algorithm in the data desensitization algorithm library in combination with a pre-set utility measurement indicator, determine the desensitization utility corresponding to each data desensitization algorithm, and based on the desensitization utility and privacy protection effect, construct and train a desensitization decision tree model with data privacy attribute characteristics and privacy protection requirements as input and the data desensitization algorithm as output, extract the privacy attribute characteristics of the data features and add them to the desensitization decision tree model, obtain the data desensitization algorithm with the highest matching degree with the data features, and use the algorithm for data desensitization;

[0062] The desensitized data features are divided into real data and integer data, the real data are encoded into integers based on the lattice cryptography system and the integers are embedded into the lattice space through homomorphic mapping to generate lattice ciphertext, for the integer data, a fully homomorphic encryption scheme is selected through a fully homomorphic encryption algorithm, a public-private key pair is generated according to the parameters of the fully homomorphic encryption scheme and homomorphic encryption is performed to obtain a fully homomorphic encrypted ciphertext, the lattice ciphertext and the fully homomorphic encrypted ciphertext are uniformly encapsulated to generate a mixed fully homomorphic ciphertext data packet and sent to the privacy protection computing platform.

[0063] The desensitizing effect refers to the validity and practicality of the data after desensitization. The desensitizing decision tree model is based on the decision tree algorithm and applies data desensitization technology, so that the decision tree can still perform accurate classification or regression analysis under the premise of protecting privacy. The privacy attribute feature refers to the sensitive attributes that need to be protected in the data. The homomorphic mapping refers to an operation correspondence between the encryption domain and the plaintext domain, so that the operations performed on the encrypted data can be directly mapped to the corresponding operations on the plaintext data. The lattice space is a discrete point set generated by a vector, with an addition and multiplication operation structure. The lattice ciphertext is an encrypted text generated based on the lattice structure. The fully homomorphic encryption ciphertext is encrypted data generated by a fully homomorphic encryption algorithm.

[0064] Get the raw data to be processed from the data source (such as a database, file system, or real-time data stream). This data may include images, time series data, and graph structure data. The acquired raw data is imported into the data feature extraction model based on deep learning. The model contains a discriminator to determine the type of input data. The discriminator divides the raw data into three categories according to the format and characteristics of the data: image data, sequence data, and graph data. For the part that the discriminator identifies as image data, the data will be sent to a pre-set convolutional neural network to extract the local features of the image through convolution operations. The filter of the convolutional layer slides on the image to generate a feature map. The feature map is downsampled through a pooling operation (such as maximum pooling or average pooling) to reduce the size of the feature map while retaining the main features. For the part that is identified by the discriminator as sequence data (such as time series, text data), the data will be sent to a recurrent neural network. Through the recurrent connection, the hidden state of each time step is updated. For the part that is identified by the discriminator as graph data (such as social network, molecular structure data), the data will be sent to a graph neural network. The feature vector of each node will be updated according to the information of the neighboring nodes. Information is transmitted between nodes to capture the relationship and pattern in the graph structure. Feature fusion is performed through weighted summation to obtain the data features.

[0065] Collect common data desensitization algorithms, such as noise addition, permutation, clustering, etc., build an algorithm library containing multiple desensitization algorithms, design a set of utility measurement indicator system, consider multiple dimensions such as data availability, integrity, consistency, etc., quantitatively evaluate the utility loss before and after data desensitization, and use the designed utility measurement indicators to quantitatively evaluate each desensitization algorithm in the algorithm library to obtain the desensitization utility value corresponding to each algorithm. Combine the desensitization utility evaluation results and the privacy protection effect of the algorithm, and label each desensitization algorithm in the algorithm library with utility and privacy protection labels. Take the privacy attribute characteristics and privacy protection requirements of the data as the input of the decision tree model, and take the data desensitization algorithm as the output of the decision tree. Based on the desensitization algorithm utility evaluation results, train the decision tree model, establish a mapping relationship between data characteristics and the optimal desensitization algorithm, extract the privacy attribute characteristics of the data for new data desensitization requirements, and input them into the decision tree model. The decision tree model selects the desensitization algorithm that best matches the data and has the best desensitization utility in the algorithm library according to the data characteristics, and uses the optimal desensitization algorithm output by the decision tree model to desensitize the data;

[0066] The desensitized data is divided into two parts: real number type data and integer type data. For real number type data, a lattice cryptographic system is used for encoding and homomorphic mapping, and the real number data is encoded into integer form. The encoded integer is embedded into the lattice space using homomorphic mapping to generate the corresponding lattice ciphertext. For integer type data, a suitable fully homomorphic encryption scheme is selected according to the value range of the data, the required calculation type, etc. The fully homomorphic encryption scheme generates a homomorphic public-private key pair, and the public key is used to homomorphically encrypt the integer data to obtain the fully homomorphic encrypted ciphertext. The lattice ciphertext and the fully homomorphic encrypted ciphertext are uniformly encapsulated to generate a mixed fully homomorphic ciphertext data packet;

[0067] The encapsulated hybrid fully homomorphic ciphertext data packet is sent to the privacy-preserving computing platform, and the computing platform receives the ciphertext data packet.

[0068] In this embodiment, by constructing an algorithm library containing multiple desensitizing algorithms and using utility measurement indicators to quantitatively evaluate the algorithms, the desensitizing utility values ​​of each algorithm are obtained. The most appropriate desensitizing algorithm can be selected for different data, and the data utility is retained to the maximum extent while protecting data privacy, thereby improving the flexibility and adaptability of data desensitization. The hybrid homomorphic encryption scheme combining the lattice cryptographic system and fully homomorphic encryption fully considers the type characteristics of the data, and selects the most appropriate homomorphic encryption technology for different types of data. While providing homomorphic computing capabilities, it also meets the requirements of different data for the performance of encryption schemes, improves encryption efficiency, and supports privacy-preserving computing of ciphertext data on third-party platforms (such as privacy-preserving computing platforms). By using the lattice cryptographic system and fully homomorphic encryption technology, the data can be directly calculated in the encrypted ciphertext state. There is no need to decrypt the data during the calculation process, and the data content will not be leaked to the third-party platform. It can effectively prevent data leakage, protect data privacy, and provide safe and reliable technical support for data sharing and collaborative analysis. In summary, this embodiment takes into account data utility and computing performance while protecting privacy, and provides a practical, efficient and secure technical solution for privacy-preserving data analysis.

[0069] In an optional embodiment,

[0070] Combined with the preset utility measurement indicators, the data utility loss of each algorithm in the data desensitization algorithm library is quantitatively evaluated as shown in the following formula:

[0071]

[0072] Among them, LOSS represents data utility loss, α represents penalty factor, wi represents the weight of the i-th performance indicator, and Perf i () represents the i-th performance evaluation index, M Yrepresents the machine learning model trained on the desensitized data D', M X represents the machine learning model trained on the original data D, β represents the balance loss weight, Bal(D′) represents the balance of the desensitized data D′, and Bal(D) represents the balance of the original data D.

[0073] In this embodiment, by comparing the changes in indicators before and after desensitization, the impact of the desensitization algorithm on the data utility can be quantitatively evaluated. By introducing balance loss, the desensitization algorithm can be prevented from excessively perturbing the data distribution and maintaining the representativeness of the data. By comparing the performance differences between the two models on the same performance indicators, the impact of the desensitization algorithm on the machine learning task can be measured. It can directly reflect the utility performance of desensitized data in practical applications and provide a reference for selecting a suitable desensitization algorithm. In summary, this embodiment can effectively improve the quality and applicability of data desensitization, and provide more reliable and practical data support for privacy-preserving data mining.

[0074] S2. After receiving the ciphertext data, the privacy-preserving computing platform constructs a privacy-preserving data preprocessing protocol based on locally sensitive hashing and secret sharing technology, performs data filtering and data dimension compression in the ciphertext domain, performs multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on verifiable random functions, generates ciphertext computing results and returns them to the data provider, and the data provider verifies the data through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts the data through a decryption algorithm to generate a plaintext computing analysis result, and filters the data through adaptive filtering theory to obtain a pure plaintext computing analysis result;

[0075] The local sensitive hashing is a hashing technology used to map similar data points into the same bucket, thereby improving the efficiency of approximate neighbor search for high-dimensional data. The secret sharing technology is a method in cryptography used to divide secret data into multiple parts and distribute them to different participants. The data dimension compression is a method for reducing the dimension of a high-dimensional data set while maintaining the main features and information of the data as much as possible. The verifiable random function is an encryption tool that can generate random numbers and provide proofs so that anyone can verify the correctness of these random numbers. The multi-party secure computing protocol is a cryptographic protocol that enables multiple parties to jointly calculate the value of a function without revealing their respective inputs. The ciphertext domain refers to the space where the encrypted data is located. The zero-knowledge proof protocol is a cryptographic protocol that allows a prover to prove to a verifier that a statement is correct without revealing any specific information. The plaintext calculation and analysis result refers to the result obtained by directly calculating and analyzing the plaintext data. The pure plaintext calculation and analysis result refers to the result obtained by calculating and analyzing the original plaintext data that has not undergone any encryption or obfuscation processing.

[0076] In an optional embodiment,

[0077] After receiving the ciphertext data, the privacy-preserving computing platform constructs a privacy-preserving data preprocessing protocol based on locally sensitive hashing and secret sharing technology, performs data filtering and data dimension compression in the ciphertext domain, and performs multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on verifiable random functions. The ciphertext computing result is generated and returned to the data provider, and the data provider verifies it through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts it through a decryption algorithm to generate a plaintext computing analysis result, and filters it through adaptive filtering theory to obtain a pure plaintext computing analysis result, including:

[0078] After receiving the ciphertext data, the privacy protection computing platform constructs a data preprocessing protocol, wherein the data preprocessing protocol includes a data filtering step and a data dimension compression step;

[0079] In the data filtering step, according to the format of the ciphertext data, multiple groups of hash function parameters are selected based on local sensitive hashing and multiple hash functions are generated, the values ​​of the hash functions are used as the hash signatures of each data point in the ciphertext data, the signature similarity between different hash signatures is calculated and compared with a preset similarity threshold, if the signature similarity is greater than the similarity threshold, the data point corresponding to the current hash signature is filtered, and the comparison is repeated until all data points are processed to obtain a set of retained data points;

[0080] In the data dimension compression step, for each element in the reserved data point set, a polynomial is constructed, the value of the polynomial is decomposed according to the number of participants, and a plurality of low-dimensional secret shares are generated and sent to the participants;

[0081] After receiving the low-dimensional secret share, the participant generates a key pair based on a verifiable random function and broadcasts the public key to other participants. For the current computing task, each participant sends its ciphertext input and public key to the privacy protection computing platform. After collecting all ciphertext inputs and public keys, the privacy protection computing platform randomly selects a participant as the computing initiator. The computing initiator generates a random seed and calculates the output and proof of the verifiable random function. The random seed, the output of the verifiable random function and the proof are sent to the privacy protection computing platform and the correctness of the output is verified using the public key. If the verification is successful, a random value is selected as the seed of the pseudo-random number generator to generate a random mask. The random mask is homomorphically operated with the ciphertext input to obtain a mask calculation result. Based on the mask calculation result, each participant decrypts using a private key and sends the decrypted masked plaintext result to the computing initiator. The ciphertext calculation result is obtained by removing the random mask and returned to the data provider.

[0082] After the data provider receives the ciphertext calculation result, the privacy-preserving computing platform generates a public reference string based on a non-interactive zero-knowledge proof scheme and sends it to each participant. At the same time, the ciphertext calculation result is used as a statement, and the random mask is used as evidence. A zero-knowledge proof is generated by a generation proof algorithm and sent to the data provider. The data provider verifies according to the verification algorithm. If the verification passes, the data provider uses a decryption method based on homomorphic encryption to decrypt the ciphertext calculation result to obtain the plaintext calculation analysis result, initialize the pre-set filter weights and calculate the filtering error, adaptively update the filter weights according to the filtering error, and filter the plaintext calculation analysis result to obtain the pure plaintext calculation analysis result.

[0083] The hash function is a function that converts an input of any length (called a message) into an output of a fixed length (called a hash value or digest). The hash signature is a digital signature technology that generates a signature of a message by using a hash function and a digital signature algorithm. The low-dimensional secret share refers to a secret sharing scheme in which a secret is divided into shares of smaller dimensions, each share containing part of the information. The random seed is an input value used to initialize a pseudo-random number generator. The proof of a verifiable random function is a proof that a certain output is generated by a specific input through a verifiable random function. The random mask is a data masking method generated by a random number, which is used to protect sensitive data. The homomorphic operation refers to an operation performed in the ciphertext domain, the result is still ciphertext, and the corresponding plaintext operation result is the same. The masked calculation result refers to the result after calculation by applying a random mask to the data. The calculation initiator refers to the party that initiates the calculation task in distributed computing or multi-party computing. The public reference string is a pre-generated random string shared by all participants and is used for the initialization setting of the protocol. The zero-knowledge proof is a cryptographic protocol. The prover can prove the authenticity of a statement to the verifier without revealing any specific information. The filtering error refers to the difference between the output processed by the filtering algorithm and the actual value in signal processing or data analysis.

[0084] In this embodiment, data filtering based on local sensitive hashing can remove redundant and invalid data points and reduce computing overhead. Polynomial decomposition realizes data dimension compression, generates low-dimensional secret shares, and reduces data transmission and storage costs. Data preprocessing lays a good data foundation for subsequent calculations and improves the overall computing performance. The calculation initiator generates a random seed, calculates the output and proof of the verifiable random function, and other participants can verify the correctness of the output, ensuring the security and verifiability of the random mask generation process, preventing tampering and destruction by malicious participants, and ensuring the credibility of privacy-preserving calculations. Each participant can only obtain his own masked plaintext results and cannot understand the data and complete calculation results of other participants, effectively preventing data leakage and privacy infringement. Zero-knowledge proof protects the privacy of the calculation process, while providing a credibility guarantee for the calculation results, and enhancing the trust of the data provider in the calculation results. In summary, this embodiment ensures the reliability, security and practicality of the privacy-preserving data mining process, and provides a credible, efficient, privacy-preserving computing environment and result output solution for data providers and participants.

[0085] In an optional embodiment,

[0086] The weight of the filter is adaptively updated according to the filtering error as shown in the following formula:

[0087]

[0088] Among them, w(n) represents the filter weight vector at time step n, w(n-1) represents the filter weight vector at time step n-1, P(n-1) represents the filter prediction error covariance matrix at time step n-1, and m r (n) represents the input vector at time step n, w T (n-1) represents the transpose of the filter weight vector at time step n-1, represents the transpose of the input vector at time step n, and λ represents the forgetting rate.

[0089] In this embodiment, the adaptive filter can be dynamically optimized according to the data characteristics and noise conditions to achieve a more accurate and reliable filtering effect. When the data mutates or becomes abnormal, the adaptive filter can quickly adjust the weights to reduce the impact of abnormal data on the filtering results and maintain the stability of the filtering performance. The forgetting rate controls the degree to which the filter forgets historical data. By adjusting the forgetting rate, the sensitivity and stability of the filter to data changes can be balanced, thereby improving the robustness of the filter. In summary, this embodiment can adapt to the changing trend of data, reduce the impact of abnormal data, and improve the accuracy and reliability of filtering results, providing an important guarantee for obtaining high-quality plaintext calculation and analysis results.

[0090] S3. Based on the pure plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network. The pure plaintext calculation and analysis results are restored through reverse learning and adversarial generation to obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate the optimal decision result. The optimal decision result is desensitized by the data desensitization algorithm, and the desensitized optimal decision result is uploaded to the privacy protection blockchain network for distributed trusted storage and sharing.

[0091] The deep generative adversarial network is a variant of the generative adversarial network, which uses deep learning technology to generate high-quality data samples. The reverse data desensitization model is a data processing method used to restore the original data or part of the original data in the desensitized data. The dual-objective optimization model is an optimization method that aims to optimize two objective functions at the same time. The multi-objective evolutionary optimization algorithm is a type of optimization method based on evolutionary algorithms that aims to optimize multiple objective functions at the same time. The optimal balance point refers to a specific solution found in multi-objective optimization that achieves the best balance between all objectives. The optimal decision result refers to the optimal decision plan or result under given objectives or constraints selected through system analysis, optimization algorithms or other decision-making methods when faced with multiple choices or decision plans. The privacy-preserving blockchain network is a blockchain technology that aims to protect user privacy and data security while maintaining the decentralization and immutability of the blockchain.

[0092] In an optional embodiment,

[0093] Based on the pure plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network. The pure plaintext calculation and analysis results are restored through reverse learning and adversarial generation to obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate the optimal decision result. The optimal decision result is desensitized by the data desensitization algorithm, and the desensitized optimal decision result is uploaded to the privacy protection blockchain network for distributed trusted storage and sharing, including:

[0094] Based on the clean plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network, wherein the data desensitization model includes a desensitizer and a restorer, the desensitizer adopts the structure of the generator in the deep generative adversarial network, and generates desensitized data by performing a deconvolution operation on the clean plaintext calculation and analysis results. For the desensitized data, the restorer adopts the structure of the discriminator in the deep generative adversarial network, restores the desensitized data through a convolution operation, outputs the restored data, repeats the iteration and adjusts the hyperparameters of the desensitizer and restorer according to the original data corresponding to the clean plaintext calculation and analysis results until convergence or reaching the preset maximum number of iterations, and uses the restored data output by the last iteration as the analysis and calculation result;

[0095] Based on the analysis and calculation results, the data provider constructs a dual-objective optimization model with privacy protection strength and decision effectiveness as the goals according to a multi-objective evolutionary optimization algorithm, adds the analysis and calculation results to the dual-objective optimization model, randomly generates an initial population based on the analysis and calculation results, performs non-dominated sorting on the individuals in the initial population, obtains the non-dominated level and crowding degree corresponding to each individual, selects multiple individuals from the initial population to form a parent population through a binary tournament selection operator, performs crossover and mutation operations on the parent population to obtain a child population, merges the child population and the initial population to generate a candidate second-generation population and performs non-dominated sorting to obtain the dominance relationship and crowding degree between individuals in the candidate second-generation population, removes the dominated individuals, and for the retained individuals, sorts them in descending order according to the crowding degree of each individual, selects the top 10 individuals to retain, obtains the second-generation population, repeats the iteration until the preset maximum number of iterations is reached, selects the individual ranked first in the last iteration as the optimal balance point, and generates the optimal decision result based on the privacy protection strength and decision effectiveness corresponding to the optimal balance point;

[0096] The optimal decision result is desensitized by the data desensitization algorithm, the desensitized optimal decision result is digitally signed and packaged and uploaded to the privacy protection blockchain network, the block is verified by the consensus mechanism and the verified block is added to the local blockchain to complete distributed trusted storage and sharing.

[0097] The desensitizer is a method or tool for protecting sensitive data by replacing real data with meaningless data (such as random numbers, pseudonyms or encrypted data) so that unauthorized users cannot identify and use the data. The restorer is a tool or method for restoring desensitized data to original data under specific conditions. The privacy protection strength refers to the degree to which privacy protection measures protect data privacy. The decision effectiveness refers to the effect and efficiency of the decision made in achieving goals, solving problems or achieving expected results. The non-dominated sorting is a method in multi-objective optimization for classifying and sorting solution sets. The crowding degree is a measurement method for evaluating the distribution density of solutions in a multi-objective optimization solution set. The binary tournament is a selection strategy for individual selection in evolutionary algorithms. The digital signature is an encryption technology for verifying the authenticity and integrity of digital messages or files. The consensus mechanism is a protocol used in a distributed system to ensure that each node in the system reaches a consensus on the data status. The distributed trusted storage is a storage system that ensures the security and reliability of data storage through a distributed architecture and encryption technology.

[0098] Based on the pure plaintext calculation and analysis results, a reverse data desensitization model is constructed by combining the data desensitization algorithm and the deep generative adversarial network. The reverse data desensitization model includes two parts: a desensitizer and a restorer. The desensitizer adopts the structure of the generator in the deep generative adversarial network, and generates desensitized data by performing a deconvolution operation on the pure plaintext calculation and analysis results. The restorer adopts the structure of the discriminator in the deep generative adversarial network, and restores the desensitized data through a convolution operation, outputs the restored data, and repeats the iterative process of the desensitizer and restorer. The hyperparameters of the desensitizer and restorer are adjusted according to the original data corresponding to the pure plaintext calculation and analysis results. The iterative process continues until the model converges or reaches the preset maximum number of iterations, and the restored data output by the last iteration is used as the analysis and calculation result;

[0099] Based on the analysis and calculation results, the data provider builds a dual-objective optimization model according to the multi-objective evolutionary optimization algorithm. The dual-objective optimization model takes privacy protection strength and decision effectiveness as optimization goals. The analysis and calculation results are added to the dual-objective optimization model as the input data for optimization. The initial population is randomly generated based on the analysis and calculation results, and the individuals in the initial population are non-dominated. The non-dominated level and crowding degree corresponding to each individual are obtained. Multiple individuals are selected from the initial population through the binary tournament selection operator to form a parent population. The parent population is crossover and mutation operations are performed to obtain a child population, and the child population is Merge with the initial population to generate a candidate second-generation population, perform non-dominated sorting on the candidate second-generation population, obtain the dominance relationship and crowding degree between individuals in the candidate second-generation population, remove the dominated individuals in the candidate second-generation population, sort the retained individuals in descending order according to the crowding degree of each individual, select the top 10 individuals to retain, and obtain a new generation population, i.e., the second-generation population. Repeat the iteration until the preset maximum number of iterations is reached. In the last iteration, select the individual ranked first as the optimal balance point, and generate the optimal decision result based on the privacy protection strength and decision effectiveness corresponding to the optimal balance point;

[0100] The optimal decision result is desensitized through the data desensitizing algorithm, and the desensitized optimal decision result is digitally signed. The signed desensitized decision result is packaged and prepared to be uploaded to the privacy-protected blockchain network. The packaged desensitized decision result is uploaded to the privacy-protected blockchain network, and the uploaded blocks are verified through the consensus mechanism. The verified blocks are added to the local blockchain to complete distributed trusted storage and sharing.

[0101] In this embodiment, by iteratively optimizing the desensitizer and the restorer, the characteristics and information of the original data can be retained to the maximum extent while protecting privacy. Through operations such as non-dominated sorting, crowding calculation, tournament selection, crossover mutation, etc., the population is generated and optimized, and finally the optimal balance between privacy protection and decision effectiveness is found. While protecting privacy, the decision results still have sufficient effectiveness and availability, providing reliable data support for subsequent decision-making applications. Based on the distributed storage and consensus mechanism of blockchain, each participant can perform credible verification and secure access to the decision results, thereby improving the security and reliability of sharing decision results. In summary, this embodiment achieves multiple goals such as privacy protection, decision effectiveness balance, secure sharing and trusted storage, provides strong technical support for the generation and application of decision results in privacy-preserving data mining, ensures the privacy security of data providers, improves the availability and reliability of decision results, promotes data sharing and collaboration among multiple parties, and lays a solid foundation for the practical application of privacy-preserving data mining.

[0102] In an optional embodiment,

[0103] The crowding degree of each individual is calculated in the non-dominated sorting process as shown in the following formula:

[0104]

[0105] Among them, d g represents the crowding degree of individual g, K represents the number of objective functions, represents the next objective function value in the non-dominated level where individual g is located on the kth objective function, represents the previous objective function value in the non-dominated level of individual g on the kth objective function, represents the maximum value of the kth objective function in the current non-dominated level, represents the minimum value of the kth objective function in the current non-dominated level, x g represents the decision variable vector corresponding to individual g, x h represents the decision variable vector corresponding to individual h, N represents the population size in the individual, and ‖‖ represents the Euclidean distance.

[0106] In this embodiment, by giving priority to retaining individuals with smaller crowding, the uniform distribution of the population in the objective function space can be promoted, the diversity of the population can be maintained, and the population can be prevented from converging to the local optimal solution too early. By giving priority to retaining individuals with smaller crowding, the aggregation of the population in the decision variable space can be promoted, the convergence of the population can be improved, and the convergence speed of the population to the global optimal solution can be accelerated. The crowding calculation formula comprehensively considers the distribution of individuals in the objective function space and the decision variable space. By weighted summation of the two terms, the exploration ability and utilization ability of the algorithm can be balanced. By balancing exploration and utilization, the global search ability and local search ability of the algorithm can be improved, and the algorithm can be prevented from falling into the local optimum. In summary, this embodiment comprehensively considers the distribution characteristics of individuals in different spaces. By giving priority to retaining individuals with smaller crowding, the uniform distribution and aggregation of the population can be promoted, and the convergence of the population to the global optimal solution can be accelerated. At the same time, the exploration ability and utilization ability of the algorithm are taken into account, which provides strong support for the performance improvement of the multi-objective optimization algorithm, enables the algorithm to solve complex optimization problems more efficiently and reliably, and provides important guarantees for the optimization of decision results in privacy-preserving data mining.

[0107] Figure 2 FIG. 1 is a schematic diagram of a privacy protection system based on big data security and privacy computing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0108] The first unit is used to obtain the original data to be processed, perform feature extraction through a data feature extraction model based on deep learning, obtain data features and select a data desensitization algorithm according to the data features, perform data desensitization through the selected data desensitization algorithm, encrypt the desensitized privacy data through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, generate homomorphically encrypted ciphertext data and send it to a privacy protection computing platform;

[0109] The second unit is used for the privacy-preserving computing platform to construct a privacy-preserving data preprocessing protocol based on local sensitive hashing and secret sharing technology after receiving the ciphertext data, perform data filtering and data dimension compression in the ciphertext domain, perform multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on a verifiable random function, generate a ciphertext computing result and return it to the data provider, and the data provider verifies it through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts it through a decryption algorithm to generate a plaintext computing analysis result, and filters it through adaptive filtering theory to obtain a pure plaintext computing analysis result;

[0110] The third unit is used to construct a reverse data desensitization model based on the pure plaintext calculation and analysis results in combination with the data desensitization algorithm and the deep generative adversarial network, restore the pure plaintext calculation and analysis results through reverse learning and adversarial generation, and obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate an optimal decision result, desensitizes the optimal decision result through the data desensitization algorithm, and uploads the desensitized optimal decision result to the privacy protection blockchain network for distributed trusted storage and sharing.

[0111] According to a third aspect of the embodiments of the present invention,

[0112] An electronic device is provided, comprising:

[0113] processor;

[0114] a memory for storing processor-executable instructions;

[0115] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0116] A fourth aspect of the embodiments of the present invention is:

[0117] 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.

[0118] 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.

[0119] 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 privacy protection method based on big data security and privacy computing, characterized in that: include: Obtaining the original data to be processed, extracting features through a data feature extraction model based on deep learning, obtaining data features and selecting a data desensitization algorithm according to the data features, desensitizing the data through the selected data desensitization algorithm, encrypting the desensitized privacy data through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, generating homomorphically encrypted ciphertext data and sending it to a privacy protection computing platform; After receiving the ciphertext data, the privacy-preserving computing platform constructs a privacy-preserving data preprocessing protocol based on locally sensitive hashing and secret sharing technology, performs data filtering and data dimension compression in the ciphertext domain, and performs multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on verifiable random functions, generates ciphertext computing results and returns them to the data provider, and the data provider verifies the data through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts the data through a decryption algorithm to generate a plaintext computing analysis result, and filters the data through adaptive filtering theory to obtain a pure plaintext computing analysis result; Based on the pure plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network, and the pure plaintext calculation and analysis results are restored through reverse learning and adversarial generation to obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate the optimal decision result. The optimal decision result is desensitized by the data desensitization algorithm, and the desensitized optimal decision result is uploaded to the privacy protection blockchain network for distributed trusted storage and sharing; The raw data to be processed is obtained, and features are extracted through a data feature extraction model based on deep learning to obtain data features and select a data desensitization algorithm according to the data features, and data desensitization is performed through the selected data desensitization algorithm. For the desensitized privacy data, the desensitized privacy data is encrypted through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, and homomorphically encrypted ciphertext data is generated and sent to the privacy protection computing platform, including: The raw data to be processed is obtained and added to a data feature extraction model based on deep learning. The data feature extraction model determines the data type through a discriminator, and divides the raw data into image data, sequence data, and graph data. For the image data, feature extraction is performed through a convolutional layer and a pooling layer of a pre-set convolutional neural network. For the sequence data, feature extraction is performed on hidden state updates through a recurrent neural network. For the graph data, feature extraction is performed through node updates and message passing in a graph neural network. The extracted features are feature fused to obtain the data features. Construct a data desensitization algorithm library containing multiple data desensitization algorithms, quantitatively evaluate the data utility loss of each algorithm in the data desensitization algorithm library in combination with a pre-set utility measurement indicator, determine the desensitization utility corresponding to each data desensitization algorithm, and based on the desensitization utility and privacy protection effect, construct and train a desensitization decision tree model with data privacy attribute characteristics and privacy protection requirements as input and the data desensitization algorithm as output, extract the privacy attribute characteristics of the data features and add them to the desensitization decision tree model, obtain the data desensitization algorithm with the highest matching degree with the data features, and use the algorithm for data desensitization; The desensitized data features are divided into real data and integer data, the real data are encoded into integers based on the lattice cryptographic system, and the integers are embedded into the lattice space through homomorphic mapping to generate lattice ciphertext, for the integer data, a fully homomorphic encryption scheme is selected through a fully homomorphic encryption algorithm, a public-private key pair is generated according to the parameters of the fully homomorphic encryption scheme, and homomorphic encryption is performed to obtain a fully homomorphic encrypted ciphertext, the lattice ciphertext and the fully homomorphic encrypted ciphertext are uniformly encapsulated, a mixed fully homomorphic ciphertext data packet is generated, and the mixed fully homomorphic ciphertext data packet is sent to the privacy protection computing platform; Combined with the preset utility measurement indicators, the data utility loss of each algorithm in the data desensitization algorithm library is quantitatively evaluated as shown in the following formula: Among them, LOSS represents data utility loss, α represents penalty factor, wi represents the weight of the i-th performance indicator, and Perf i () represents the i-th performance evaluation index, M Y represents the machine learning model trained on the desensitized data D', M X represents the machine learning model trained on the original data D, β represents the balance loss weight, Bal(D′) represents the balance of the desensitized data D′, and Bal(D) represents the balance of the original data D.

2. The method according to claim 1, characterized in that After receiving the ciphertext data, the privacy-preserving computing platform constructs a privacy-preserving data preprocessing protocol based on locally sensitive hashing and secret sharing technology, performs data filtering and data dimension compression in the ciphertext domain, and performs multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on verifiable random functions. The ciphertext computing result is generated and returned to the data provider, and the data provider verifies it through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts it through a decryption algorithm to generate a plaintext computing analysis result, and filters it through adaptive filtering theory to obtain a pure plaintext computing analysis result, including: After receiving the ciphertext data, the privacy protection computing platform constructs a data preprocessing protocol, wherein the data preprocessing protocol includes a data filtering step and a data dimension compression step; In the data filtering step, according to the format of the ciphertext data, multiple groups of hash function parameters are selected based on local sensitive hashing and multiple hash functions are generated, the values ​​of the hash functions are used as the hash signatures of each data point in the ciphertext data, the signature similarity between different hash signatures is calculated and compared with a preset similarity threshold, if the signature similarity is greater than the similarity threshold, the data point corresponding to the current hash signature is filtered, and the comparison is repeated until all data points are processed to obtain a set of retained data points; In the data dimension compression step, for each element in the reserved data point set, a polynomial is constructed, the value of the polynomial is decomposed according to the number of participants, and a plurality of low-dimensional secret shares are generated and sent to the participants; After receiving the low-dimensional secret share, the participant generates a key pair based on a verifiable random function and broadcasts the public key to other participants. For the current computing task, each participant sends its ciphertext input and public key to the privacy protection computing platform. After collecting all ciphertext inputs and public keys, the privacy protection computing platform randomly selects a participant as the computing initiator. The computing initiator generates a random seed and calculates the output and proof of the verifiable random function. The random seed, the output of the verifiable random function and the proof are sent to the privacy protection computing platform and the correctness of the output is verified using the public key. If the verification is successful, a random value is selected as the seed of the pseudo-random number generator to generate a random mask. The random mask is homomorphically operated with the ciphertext input to obtain a mask calculation result. Based on the mask calculation result, each participant decrypts using a private key and sends the decrypted masked plaintext result to the computing initiator. The ciphertext calculation result is obtained by removing the random mask and returned to the data provider. After the data provider receives the ciphertext calculation result, the privacy-preserving computing platform generates a public reference string based on a non-interactive zero-knowledge proof scheme and sends it to each participant. At the same time, the ciphertext calculation result is used as a statement, and the random mask is used as evidence. A zero-knowledge proof is generated by a generation proof algorithm and sent to the data provider. The data provider verifies according to the verification algorithm. If the verification passes, the data provider uses a decryption method based on homomorphic encryption to decrypt the ciphertext calculation result to obtain the plaintext calculation analysis result, initialize the pre-set filter weights and calculate the filtering error, adaptively update the filter weights according to the filtering error, and filter the plaintext calculation analysis result to obtain the pure plaintext calculation analysis result.

3. The method according to claim 2, characterized in that The weight of the filter is adaptively updated according to the filtering error as shown in the following formula: Among them, w(n) represents the filter weight vector at time step n, w(n-1) represents the filter weight vector at time step n-1, P(n-1) represents the filter prediction error covariance matrix at time step n-1, and m r (n) represents the input vector at time step n, w T (n-1) represents the transpose of the filter weight vector at time step n-1, represents the transpose of the input vector at time step n, and λ represents the forgetting rate.

4. The method according to claim 1, characterized in that: Based on the pure plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network. The pure plaintext calculation and analysis results are restored through reverse learning and adversarial generation to obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate the optimal decision result. The optimal decision result is desensitized by the data desensitization algorithm, and the desensitized optimal decision result is uploaded to the privacy protection blockchain network for distributed trusted storage and sharing, including: Based on the clean plaintext calculation and analysis results, a reverse data desensitization model is constructed in combination with the data desensitization algorithm and the deep generative adversarial network, wherein the data desensitization model includes a desensitizer and a restorer, the desensitizer adopts the structure of the generator in the deep generative adversarial network, and generates desensitized data by performing a deconvolution operation on the clean plaintext calculation and analysis results. For the desensitized data, the restorer adopts the structure of the discriminator in the deep generative adversarial network, restores the desensitized data through a convolution operation, outputs the restored data, repeats the iteration and adjusts the hyperparameters of the desensitizer and restorer according to the original data corresponding to the clean plaintext calculation and analysis results until convergence or reaching the preset maximum number of iterations, and uses the restored data output by the last iteration as the analysis and calculation result; Based on the analysis and calculation results, the data provider constructs a dual-objective optimization model with privacy protection strength and decision effectiveness as the goals according to a multi-objective evolutionary optimization algorithm, adds the analysis and calculation results to the dual-objective optimization model, randomly generates an initial population based on the analysis and calculation results, performs non-dominated sorting on the individuals in the initial population, obtains the non-dominated level and crowding degree corresponding to each individual, selects multiple individuals from the initial population to form a parent population through a binary tournament selection operator, performs crossover and mutation operations on the parent population to obtain a child population, merges the child population and the initial population to generate a candidate second-generation population and performs non-dominated sorting to obtain the dominance relationship and crowding degree between individuals in the candidate second-generation population, removes the dominated individuals, and for the retained individuals, sorts them in descending order according to the crowding degree of each individual, selects the top 10 individuals to retain, obtains the second-generation population, repeats the iteration until the preset maximum number of iterations is reached, selects the individual ranked first in the last iteration as the optimal balance point, and generates the optimal decision result based on the privacy protection strength and decision effectiveness corresponding to the optimal balance point; The optimal decision result is desensitized by the data desensitization algorithm, the desensitized optimal decision result is digitally signed and packaged and uploaded to the privacy protection blockchain network, the block is verified by the consensus mechanism and the verified block is added to the local blockchain to complete distributed trusted storage and sharing.

5. The method according to claim 4, characterized in that The crowding degree of each individual is calculated in the non-dominated sorting process as shown in the following formula: Among them, d g represents the crowding degree of individual g, K represents the number of objective functions, represents the next objective function value in the non-dominated level where individual g is located on the kth objective function, represents the previous objective function value in the non-dominated level of individual g on the kth objective function, represents the maximum value of the kth objective function in the current non-dominated level, represents the minimum value of the kth objective function in the current non-dominated level, x g represents the decision variable vector corresponding to individual g, x h represents the decision variable vector corresponding to individual h, N represents the population size in the individual, and ‖‖ represents the Euclidean distance.

6. A privacy protection system based on big data security and privacy computing, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to obtain the original data to be processed, perform feature extraction through a data feature extraction model based on deep learning, obtain data features and select a data desensitization algorithm according to the data features, perform data desensitization through the selected data desensitization algorithm, encrypt the desensitized privacy data through a hybrid fully homomorphic encryption algorithm based on a lattice cryptographic system and fully homomorphic encryption, generate homomorphically encrypted ciphertext data and send it to a privacy protection computing platform; The second unit is used for the privacy-preserving computing platform to construct a privacy-preserving data preprocessing protocol based on local sensitive hashing and secret sharing technology after receiving the ciphertext data, perform data filtering and data dimension compression in the ciphertext domain, perform multi-party secure computing in the ciphertext domain through a secure multi-party computing protocol based on a verifiable random function, generate a ciphertext computing result and return it to the data provider, and the data provider verifies it through a zero-knowledge proof protocol. If the verification is successful, the data provider decrypts it through a decryption algorithm to generate a plaintext computing analysis result, and filters it through adaptive filtering theory to obtain a pure plaintext computing analysis result; The third unit is used to construct a reverse data desensitization model based on the pure plaintext calculation and analysis results in combination with the data desensitization algorithm and the deep generative adversarial network, restore the pure plaintext calculation and analysis results through reverse learning and adversarial generation, and obtain the analysis and calculation results corresponding to the original data. Based on the analysis and calculation results, the data provider constructs and solves a dual-objective optimization model based on a multi-objective evolutionary optimization algorithm to obtain the optimal balance point and generate an optimal decision result, desensitizes the optimal decision result through the data desensitization algorithm, and uploads the desensitized optimal decision result to the privacy protection blockchain network for distributed trusted storage and sharing.

7. 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 5.

8. 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 5 is implemented.

Images