Data anonymous tracing method and device, model training method and device, equipment, medium and product
By using secret-shared fragments obtained through anonymous privacy intersection as input in the data anonymization learning model, the problem of low quality in sensitive data anonymization is solved, achieving efficient data privacy protection and model training integration, and improving the accuracy and security of model training.
Patent Information
- Application Number
- CN202511707761.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-20
- Publication Date
- 2026-02-10
AI Technical Summary
Existing methods for concealing sensitive data have low concealment quality in scenarios with large differences in data feature distribution, affecting the accuracy of model training and increasing computational and communication overhead.
By using the secret shared fragments of the intersection data obtained through anonymous privacy intersection as input, and training the data anonymity learning model, we can ensure that the data flows in encrypted or fragmented form at each stage, avoiding the direct use of the original data.
It achieves full anonymity of sensitive information, improves the anonymity quality of model training, reduces computational and communication overhead, and ensures the security of data privacy.
Smart Images
Figure CN121502811A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] Embodiments of the present application relate to the technical field of data security, and in particular to a data privacy tracking method and device, a model training method and device, equipment, media and products. BACKGROUND
[0002] With the increasingly close contact between various industries, there is an increasing demand for data cooperation in high-sensitive industries, but at the same time, both parties to the cooperation are concerned about the leakage of sensitive information. At present, the industry uses a private set union (PSU) protocol to securely align the data of each participant, thereby avoiding the leakage of intersection element information by traditional private set intersection (PSI) algorithms. The private set union method uses a synthetic data generation strategy to generate synthetic labels and synthetic features for non-intersection data, so that both parties can effectively utilize data while protecting their own data privacy when jointly training a model. However, in scenarios where the distribution of data features is significantly different, the quality of the synthetic data generated by this scheme may not be guaranteed, which in turn affects the accuracy of subsequent model training and makes it impossible to achieve high-quality full privacy tracking of sensitive information data. In addition, training using union data requires processing more redundant samples, significantly increasing the computational and communication overhead. SUMMARY
[0003] Embodiments of the present application provide a data privacy tracking method and device, a model training method and device, equipment, media and products to solve the problem of low privacy tracking quality in existing sensitive data privacy tracking methods.
[0004] To solve the above technical problems, the present application is implemented as follows:
[0005] In a first aspect, embodiments of the present application provide a training method for a data privacy tracking learning model, the data privacy tracking learning model being applied to at least one first computing participant and at least one second computing participant, and the training method comprising:
[0006] The first computing participant obtains target feature data shared shards and target label data shared shards, wherein the target feature data shared shards and the target label data shared shards are obtained by extracting shared shards corresponding to target feature data and target label data in intersection first computing participant shared shards and intersection second computing participant shared shards, and the intersection first computing participant shared shards and the intersection second computing participant shared shards are obtained from secret shared shards obtained by performing intersection privacy set intersection on elements having the same unique identifier in a first computing participant dataset and a second computing participant dataset according to associated data;
[0007] The first computational participant inputs the target feature data shared slice into the data anonymity learning model to be trained. The data anonymity learning model to be trained obtains the shared slice corresponding to the predicted label data based on the target feature data shared slice and the initial model parameters of the first computational participant. The model parameters of the data anonymity learning model are then optimized based on the shared slice corresponding to the predicted label data, the target label data shared slice, and the loss function to obtain the trained anonymity learning model.
[0008] Optionally, the data anonymity learning model to be trained obtains the shared slices corresponding to the predicted label data based on the shared slices of the target feature data and the initial model parameters of the first computational participant, and optimizes the model parameters of the data anonymity learning model based on the shared slices corresponding to the predicted label data, the shared slices of the target label data, and the loss function to obtain the trained anonymity learning model, including:
[0009] The first computational participant trains the data anonymity learning model to be trained using a logistic regression model, specifically including:
[0010] The first computational participant obtains initial model parameters through the logistic regression model and obtains the model parameter sharing slices corresponding to the initial model parameters; divides the target feature data sharing slices into multiple batches; for each batch of the target feature data sharing slices, a computation process is executed, the computation process including: obtaining the shared slices corresponding to the predicted label data based on the target feature data sharing slices and the model parameter sharing slices; calculating the gradient of the loss function with respect to the model parameters based on the shared slices corresponding to the predicted label data and the target label data sharing slices, performing gradient descent, and updating the model parameters, until the computation process has been executed for each batch of the target feature data sharing slices and a preset number of iterations has been reached, outputting the final model parameter sharing slices, and obtaining the trained stealth learning model based on the final model parameter sharing slices.
[0011] Optionally, the data anonymity learning model to be trained obtains the shared slices corresponding to the predicted label data based on the shared slices of the target feature data and the initial model parameters of the first computational participant, and optimizes the model parameters of the data anonymity learning model based on the shared slices corresponding to the predicted label data, the shared slices of the target label data, and the loss function to obtain the trained anonymity learning model, including:
[0012] The first computational participant trains the data anonymity learning model to be trained using a support vector machine model, specifically including:
[0013] The first computational participant obtains initial model parameters through a support vector machine (SVM) model and obtains a shared model parameter slice corresponding to the initial model parameters. A target batch is obtained by randomly sampling from the shared target feature data slice. The shared target feature data slice containing elements in the target batch that do not satisfy the SVM constraints is extracted, and a computation process is executed. This computation process includes: obtaining a shared slice corresponding to the predicted label data based on the shared target feature data slice and the shared model parameter slice; setting the learning rate of the SVM model; updating the model parameters based on the learning rate, the shared target feature data slice, the shared slice corresponding to the predicted label data, and the shared target label data slice until a preset iteration round is reached; outputting the final shared model parameter slice; and obtaining the trained stealth learning model based on the final shared model parameter slice.
[0014] Optionally, obtaining the shared fragment corresponding to the predicted label data based on the shared fragment of the target feature data and the initial model parameters of the first computational participant includes:
[0015] The shared fragments corresponding to the predicted label data are obtained using a secure multi-party computation protocol based on the target feature data shared fragments and the initial model parameters of the first computation participant. The secure multi-party computation protocol includes at least one of the following: an addition protocol, a multiplication protocol, and a numerical comparison protocol.
[0016] Optionally, when the first computing participant is the first computing participant, the first computing participant acquires the target feature data sharing fragment and the target label data sharing fragment, including:
[0017] Obtain a first computational participant dataset, wherein each element in the first computational participant dataset includes: a unique identifier and associated data;
[0018] The unique identifier of the first computing participant's dataset is processed using a hash function to obtain a first hash value. The first hash value is then multiplied and the associated data is encrypted to obtain a first encrypted dataset. After the first encrypted dataset is randomly rearranged, it is sent to the second computing participant.
[0019] The system receives a second encrypted dataset sent by the second computing participant and a shared fragment ciphertext of the associated data of the first computing participant. The second encrypted dataset is obtained by the second computing participant processing the unique identifier of the dataset using a hash function to obtain a second hash value, multiplying the second hash value, and then encrypting the associated data. The shared fragment ciphertext of the associated data of the first computing participant is obtained by multiplying the first hash value in the first encrypted dataset, encrypting a first random number, subtracting the encrypted first random number from the associated data of the first encrypted dataset, and finally randomly rearranging the data.
[0020] The second hash value in the second encrypted dataset is exponentially multiplied to obtain the shared fragmented ciphertext of the data associated with the second computation participants.
[0021] Extract the shared fragments of the associated data of elements with equal hash values from the shared fragment ciphertext of the first computation participant's associated data and the shared fragment ciphertext of the second computation participant's associated data to obtain the first intersection;
[0022] Encrypt a second random number, subtract the encrypted second random number from the associated data of the second computing participant in the first intersection to obtain the shared fragment ciphertext of the associated data of the second computing participant in the intersection, and send the shared fragment ciphertext of the associated data of the second computing participant to the second computing participant. The second computing participant decrypts the shared fragment ciphertext of the associated data of the second computing participant and merges the decrypted shared fragment ciphertext of the associated data of the second computing participant with the first random number to obtain the shared fragment of the second computing participant in the intersection.
[0023] Based on the elements in the first intersection, the shared fragment ciphertext of the first computational participant's associated data is decrypted, and the decrypted shared fragment ciphertext of the first computational participant's associated data is merged with the second random number to obtain the first computational participant's shared fragment of the intersection.
[0024] Optionally, the first random number and the second random number are uniformly and randomly selected from the plaintext space of the homomorphic encryption framework agreed upon by the first computing participant and the second computing participant.
[0025] Secondly, embodiments of the present invention provide a data anonymity learning method, including:
[0026] Obtain feature data from the target dataset;
[0027] The target shared fragments are obtained based on the feature data of the target dataset;
[0028] The target shared fragment is input into the data anonymity learning model, and the data anonymity learning model outputs the shared fragment of the label corresponding to the feature data based on the target shared fragment.
[0029] Thirdly, embodiments of the present invention provide a training apparatus for a data anonymity learning model, comprising:
[0030] The first acquisition module is used by the first computing participant to acquire the target feature data sharing fragment and the target label data sharing fragment. The target feature data sharing fragment and the target label data sharing fragment are obtained by extracting the sharing fragments corresponding to the target feature data and the target label data from the intersection of the first computing participant sharing fragment and the intersection of the second computing participant sharing fragment. The intersection of the first computing participant sharing fragment and the intersection of the second computing participant sharing fragment are obtained from the secret sharing fragment obtained after performing anonymity and privacy calculation on the associated data of elements with the same unique identifier in the first computing participant dataset and the second computing participant dataset.
[0031] The first processing module is used to input the target feature data shared slice into the data anonymity learning model to be trained. The data anonymity learning model to be trained obtains the shared slice corresponding to the predicted label data based on the target feature data shared slice and the initial model parameters of the first computing participant. The model parameters of the data anonymity learning model are optimized based on the shared slice corresponding to the predicted label data, the target label data shared slice, and the loss function to obtain the trained anonymity learning model.
[0032] Fourthly, embodiments of the present invention provide a data anonymity learning device, comprising:
[0033] The second acquisition module is used to acquire feature data of the target dataset;
[0034] The second processing module is used to obtain the target shared fragments based on the feature data of the target dataset.
[0035] The third processing module is used to input the target shared fragment into the data anonymity learning model, and the data anonymity learning model outputs the shared fragment of the label corresponding to the feature data based on the target shared fragment.
[0036] Fifthly, embodiments of the present invention provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the training method of the data stealth learning model as described in any one of the first aspects, or the steps in the data stealth learning method described in the second aspect.
[0037] In a sixth aspect, embodiments of the present invention provide a readable storage medium storing a program or instructions that, when executed by a processor, implement the training method of the data-stealth learning model as described in any one of the first aspects, or the steps in the data-stealth learning method described in the second aspect.
[0038] In a seventh aspect, embodiments of the present invention provide a computer program product, including computer instructions that, when executed by a processor, implement a training method for a data-stealth learning model as described in any one of the first aspects, or the steps in a data-stealth learning method as described in the second aspect.
[0039] In this invention, a first computational participant acquires a shared fragment of target feature data and a shared fragment of target label data. These shared fragments are obtained by extracting the shared fragments corresponding to the target feature data and target label data from the intersection of the first and second computational participant shared fragments. The first and second shared fragments are obtained from a secret shared fragment obtained by performing anonymity and privacy intersection calculations on the associated data of elements with the same unique identifier in the first and second computational participant datasets. The first computational participant inputs the shared fragment of target feature data into a data anonymity learning model to be trained. The data anonymity learning model to be trained obtains the shared fragments corresponding to the predicted label data based on the shared fragments of target feature data and the initial model parameters of the first computational participant. The model parameters of the data anonymity learning model are then optimized based on the shared fragments corresponding to the predicted label data, the shared fragments of target label data, and a loss function to obtain the trained anonymity learning model. By directly using the secret-shared fragments of the intersection data obtained from the anonymous privacy intersection as input for anonymous learning, seamless integration of data alignment and model training is achieved. This ensures that data flows in encrypted or fragmented form at each stage, guaranteeing the complete anonymity of sensitive information and solving the problem of low anonymity quality in existing sensitive data anonymization methods. Attached Figure Description
[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:
[0041] Figure 1 This is a flowchart of a training method for a data anonymity learning model provided in an embodiment of the present invention;
[0042] Figure 2 This is a flowchart of the privacy intersection stage of a training method for a data anonymity learning model provided in an embodiment of the present invention;
[0043] Figure 3 This is a flowchart illustrating the training method for a data anonymity learning model provided in this embodiment of the invention, specifically for training a multi-class logistic regression model.
[0044] Figure 4 This is a flowchart illustrating the training method for a data anonymity learning model provided in an embodiment of the present invention, specifically a training method for a multi-class support vector machine model.
[0045] Figure 5 This is a flowchart of the multiplication protocol for a training method of a data anonymity learning model provided in an embodiment of the present invention;
[0046] Figure 6 This is a flowchart of a numerical comparison protocol for a training method of a data anonymity learning model provided in an embodiment of the present invention;
[0047] Figure 7 This is a flowchart of a data anonymity learning method provided in an embodiment of the present invention;
[0048] Figure 8 This is a schematic diagram of a training device for a data anonymity learning model provided in an embodiment of the present invention;
[0049] Figure 9 This is a schematic diagram of a data anonymization learning device provided in an embodiment of the present invention;
[0050] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0052] Please refer to Figure 1 This invention provides a method for training a data anonymity learning model, wherein the data anonymity learning model is applied to at least one first computational participant and at least one second computational participant, and the training method includes:
[0053] Step 11: The first computational participant obtains the target feature data shared fragment and the target label data shared fragment. The target feature data shared fragment and the target label data shared fragment are obtained by extracting the shared fragments corresponding to the target feature data and the target label data from the intersection of the first computational participant shared fragment and the intersection of the second computational participant shared fragment. The intersection of the first computational participant shared fragment and the intersection of the second computational participant shared fragment are obtained from the secret shared fragment obtained after performing anonymity and privacy calculation on the associated data of elements with the same unique identifier in the first computational participant dataset and the second computational participant dataset.
[0054] In this embodiment of the invention, the first shared fragment of the intersection and the second shared fragment of the intersection are obtained by processing elements with the same unique identifier in the first and second datasets of the computational participants using a hash function, aligning them using dual DH key exchange technology, and using homomorphic encryption to protect and fragment the associated data. This achieves the intersection, alignment, and secret sharing of the datasets.
[0055] In this embodiment of the invention, optionally, the first computing participant acquires the target feature data sharing fragment and the target label data sharing fragment, including:
[0056] Obtain a first computational participant dataset, wherein each element in the first computational participant dataset includes: a unique identifier and associated data;
[0057] The unique identifier of the first computing participant's dataset is processed using a hash function to obtain a first hash value. The first hash value is then multiplied and the associated data is encrypted to obtain a first encrypted dataset. After the first encrypted dataset is randomly rearranged, it is sent to the second computing participant.
[0058] The system receives a second encrypted dataset sent by the second computing participant and a shared fragment ciphertext of the associated data of the first computing participant. The second encrypted dataset is obtained by the second computing participant processing the unique identifier of the dataset using a hash function to obtain a second hash value, multiplying the second hash value, and then encrypting the associated data. The shared fragment ciphertext of the associated data of the first computing participant is obtained by multiplying the first hash value in the first encrypted dataset, encrypting a first random number, subtracting the encrypted first random number from the associated data of the first encrypted dataset, and finally randomly rearranging the data.
[0059] The second hash value in the second encrypted dataset is exponentially multiplied to obtain the shared fragmented ciphertext of the data associated with the second computation participants.
[0060] Extract the shared fragments of the associated data of elements with equal hash values from the shared fragment ciphertext of the first computation participant's associated data and the shared fragment ciphertext of the second computation participant's associated data to obtain the first intersection;
[0061] Encrypt a second random number, subtract the encrypted second random number from the associated data of the second computing participant in the first intersection to obtain the shared fragment ciphertext of the associated data of the second computing participant in the intersection, and send the shared fragment ciphertext of the associated data of the second computing participant to the second computing participant. The second computing participant decrypts the shared fragment ciphertext of the associated data of the second computing participant and merges the decrypted shared fragment ciphertext of the associated data of the second computing participant with the first random number to obtain the shared fragment of the second computing participant in the intersection.
[0062] Based on the elements in the first intersection, the shared fragment ciphertext of the first computational participant's associated data is decrypted, and the decrypted shared fragment ciphertext of the first computational participant's associated data is merged with the second random number to obtain the first computational participant's shared fragment of the intersection.
[0063] In this embodiment of the invention, each element in the first computational participant dataset includes: a unique identifier and associated data. The associated data consists of features other than the unique identifier. Before performing the privacy intersection operation, the computational participants agree on a public-key encryption framework that satisfies additive homomorphism and a hash function H, thereby mapping the unique identifiers of the elements in the computational participant dataset to a cyclic group that satisfies the DH (Decisional Diffie-Hellman, DDH) difficulty assumption. The first computational participant can be a first computational participant, and the second computational participant can be a second computational participant. Please refer to [reference needed]. Figure 2 Specifically:
[0064] The first participant in the computation runs a key generation algorithm to obtain its private key. The public key of the first computation participant And choose an order less than that of the cyclic group. random numbers The hash key serves as the first computational participant;
[0065] The first computing participant's dataset Each element in Calculate the hash value of the unique identifier (ID) for each element. And calculate the power of the hash value. ,right Related data use Encryption yields ciphertext At this point, the first encrypted dataset from the first computational participant is obtained. .
[0066] The first computing participant will use the first encrypted dataset After random rearrangement, and Send to the second computing participant.
[0067] The second computation participant runs a key generation algorithm to obtain their private key. The public key of the second computation participant And choose an order less than that of the cyclic group. random numbers As the hash key for the second computational participant;
[0068] The second computational participant has a dataset for the second computational participant. Each element Calculate the hash value of the unique identifier (ID) for each element. And calculate the power of the hash value. Its associated data use Encryption yields ciphertext At this point, the second encrypted dataset from the second computational participant is obtained. .
[0069] The second calculation participant Random rearrangement;
[0070] The second calculation participant calculate The second calculation participant is each Generate the first random number The second computational participant, as the first computational participant, shares the sharded data associated with it and uses it. encryption ,calculate The first computational participant obtains the shared fragment ciphertext of the data associated with the first computational participant. .
[0071] The second calculation participants are randomly arranged. Rearrangement and will , and Send back to the first calculation participant.
[0072] The first calculation participant calculates ,right , obtain the set .
[0073] The first calculation participant selects all those that meet the criteria. elements , to obtain intersection ;
[0074] The first computing participant is Related data of the second computing participant Generate a second random number As the first computational participant sharing the data associated with the second computational participant in the intersection, it uses... encryption get ,calculate The second computation participant shared fragment ciphertext of the intersection of the second computation participant's associated data. .
[0075] For each intersection element The first computing participant will use Decryption reveals that the first computational participant sharing the intersection partitions. .
[0076] The first computing participant will Send it back to the second calculation participant.
[0077] The second calculation participant uses Decryption, for and utilize inverse function Index to The first computational participant shares fragmented ciphertext with the data associated with it. The second computational participant that obtains the intersection shares the fragment. .
[0078] In this embodiment of the invention, the unique identifier of the data is processed using a hash function during the anonymous privacy intersection stage, and aligned using a dual key exchange technique, thereby completing the protection and fragmentation of the associated data and realizing the intersection, alignment, and secret sharing of the dataset.
[0079] In this embodiment of the invention, optionally, the first random number and the second random number are uniformly and randomly selected from the plaintext space of the homomorphic encryption framework agreed upon by the first computing participant and the second computing participant.
[0080] In this embodiment of the invention, uniformly and randomly selecting random numbers in the plaintext space of the homomorphic encryption framework can effectively avoid predictability, reduce the risk of being exploited by attackers, and thus enhance the security of data protection. Furthermore, the selection of the first and second random numbers conforms to the requirements of homomorphic encryption. Under the homomorphic encryption framework, these random numbers can be used for encrypted calculations, ensuring that the correctness of the final result is not affected when performing addition or multiplication operations. More complex functions can be implemented in homomorphic calculations, and both parties to the calculation can effectively perform necessary encryption and decryption operations, which also enhances the versatility and efficiency of the system.
[0081] Step 12: The first computational participant inputs the target feature data shared slice into the data anonymity learning model to be trained. The data anonymity learning model to be trained obtains the shared slice corresponding to the predicted label data based on the target feature data shared slice and the initial model parameters of the first computational participant. The model parameters of the data anonymity learning model are then optimized based on the shared slice corresponding to the predicted label data, the target label data shared slice, and the loss function to obtain the trained anonymity learning model.
[0082] In this embodiment of the invention, a secure multi-party computation protocol is applied to the forward propagation and parameter update steps of logistic regression and support vector machine training to train the data-hidden learning model. The training data and model parameters participate in the computation in the form of shared fragments throughout the process. The original data is not used directly, but is computed in the form of shared fragments to ensure that the private information of the participants is not leaked.
[0083] In this embodiment of the invention, a first computational participant obtains a shared fragment of target feature data and a shared fragment of target label data. These shared fragments are obtained by extracting the shared fragments corresponding to the target feature data and target label data from the intersection of the first and second computational participant shared fragments. The intersection of the first and second computational participant shared fragments is obtained from a secret shared fragment obtained after performing anonymity and privacy calculation on the associated data of elements with the same unique identifier in the first and second computational participant datasets. The first computational participant consists of the first computational participant and the second computational participant. In the anonymous privacy intersection stage, the unique identifiers of the data are processed using a hash function, aligned using dual-key exchange technology, and protected and fragmented using homomorphic encryption, thus achieving intersection, alignment, and secret sharing of the dataset. The first computational participant inputs the target feature data shared fragment into the data anonymity learning model to be trained. The data anonymity learning model to be trained obtains the shared fragment corresponding to the predicted label data based on the target feature data shared fragment and the initial model parameters of the first computational participant. The model parameters of the data anonymity learning model are then optimized based on the shared fragment corresponding to the predicted label data, the target label data shared fragment, and the loss function to obtain the trained anonymity learning model. By directly using the secret shared fragment of the intersection data obtained from the anonymous privacy intersection as the input to anonymity learning, seamless integration of data alignment and model training is achieved, ensuring that data flows in ciphertext or fragmented form at each stage, guaranteeing the end-to-end anonymity of sensitive information, and solving the problem of low anonymity quality in existing sensitive data anonymization methods.
[0084] In this embodiment of the invention, optionally, the data anonymity learning model to be trained obtains the shared slices corresponding to the predicted label data based on the target feature data shared slices and the initial model parameters of the first computational participant, and optimizes the model parameters of the data anonymity learning model based on the shared slices corresponding to the predicted label data, the target label data shared slices, and the loss function to obtain the trained anonymity learning model, including:
[0085] The first computational participant trains the data anonymity learning model to be trained using a logistic regression model, specifically including:
[0086] The first computational participant obtains initial model parameters through the logistic regression model and obtains the model parameter sharing slices corresponding to the initial model parameters; divides the target feature data sharing slices into multiple batches; for each batch of the target feature data sharing slices, a computation process is executed, the computation process including: obtaining the shared slices corresponding to the predicted label data based on the target feature data sharing slices and the model parameter sharing slices; calculating the gradient of the loss function with respect to the model parameters based on the shared slices corresponding to the predicted label data and the target label data sharing slices, performing gradient descent, and updating the model parameters, until the computation process has been executed for each batch of the target feature data sharing slices and a preset number of iterations has been reached, outputting the final model parameter sharing slices, and obtaining the trained stealth learning model based on the final model parameter sharing slices.
[0087] Please refer to Figure 3 In this embodiment of the invention, the data anonymity learning model to be trained is trained by training a multi-class logistic regression model. Specifically:
[0088] It should be noted that the binary classification features in both the first and second computational participant datasets were 0-1 encoded, and the multi-class features were one-hot encoded. Feature data was extracted from the intersection of the first and second computational participant shared fragments. and tags Shared shards That is, the target feature data sharing fragment and the target label data sharing fragment are used as training data, and the learning rate is set to 1. The specific steps include:
[0089] Initialize the model parameters using the zero matrix to obtain the initial model parameters. ,Will Secret sharing: The first and second computational participants receive shared fragments of model parameters. ;
[0090] In each iteration, the dataset Divided into several batches, each batch Includes feature data and tags [ Both parties used the multiplication agreement to calculate ( and Multiplying by matrix multiplication, since and All are secretly shared, and their product is also secretly shared ([] indicates that the result is in a secretly shared form), and the participating parties each obtain [ Shared shards and .
[0091] Calculate the shared fragments of the predicted values, that is, the shared fragments corresponding to the predicted label data. The two participating parties use a numerical comparison protocol, based on... Each element in The interval is determined Each element The value is obtained, specifically: if ,So ;like ,So ;otherwise .
[0092] The calculation formula is expressed as follows:
[0093] ;
[0094] Therefore, the shared fragments corresponding to the predicted label data obtained by both participating parties are calculated. ;
[0095] Both parties use a multiplication protocol to calculate the loss function. for gradient: ;
[0096] Gradient descent is then performed, and both sides update their parameters: ;
[0097] Repeat the above steps until the specified number of iterations is reached, then output the shared fragments of the model parameters. .
[0098] In this embodiment of the invention, the calculation process for each batch enables the model to quickly generate predicted label data sharing fragments based on target feature data and initial model parameters in each iteration cycle, accelerating the model's response speed and providing immediate feedback for subsequent gradient calculation and model updates. The gradient of the loss function is calculated based on the predicted label data sharing fragments and the target label data sharing fragments. The dynamic optimization method based on gradient descent allows the model to iteratively adjust parameters to achieve better performance.
[0099] In this embodiment of the invention, optionally, the data anonymity learning model to be trained obtains the shared slices corresponding to the predicted label data based on the target feature data shared slices and the initial model parameters of the first computational participant, and optimizes the model parameters of the data anonymity learning model based on the shared slices corresponding to the predicted label data, the target label data shared slices, and the loss function to obtain the trained anonymity learning model, including:
[0100] The first computational participant trains the data anonymity learning model to be trained using a support vector machine model, specifically including:
[0101] The first computational participant obtains initial model parameters through a support vector machine (SVM) model and obtains a shared model parameter slice corresponding to the initial model parameters. A target batch is obtained by randomly sampling from the shared target feature data slice. The shared target feature data slice containing elements in the target batch that do not satisfy the SVM constraints is extracted, and a computation process is executed. This computation process includes: obtaining a shared slice corresponding to the predicted label data based on the shared target feature data slice and the shared model parameter slice; setting the learning rate of the SVM model; updating the model parameters based on the learning rate, the shared target feature data slice, the shared slice corresponding to the predicted label data, and the shared target label data slice until a preset iteration round is reached; outputting the final shared model parameter slice; and obtaining the trained stealth learning model based on the final shared model parameter slice.
[0102] Please refer to Figure 4 In this embodiment of the invention, the data anonymity learning model to be trained is trained by training a multi-class support vector machine model. Specifically:
[0103] It should be noted that the binary classification features in both the first and second computational participant datasets were 0-1 encoded, and the multi-class features were one-hot encoded. After encoding, the values encoded as 0 in the target label data Y were replaced with -1. Feature data was extracted from the intersection of the first computational participant shared fragment and the intersection of the second computational participant shared fragment. and tags Shared shards That is, the target feature data sharing fragment and the target label data sharing fragment are used as training data, and the learning rate is set to 1. Both parties agreed on a hyperparameter Assuming have Each category. Specific steps include:
[0104] Initialize the model parameters using the zero matrix to obtain the initial model parameters. ;Will Secret sharing: The first and second computational participants receive shared fragments of model parameters. ;
[0105] In the iteration Rounds are obtained by randomly sampling from the dataset. Each batch Includes feature data and tags [ Both parties used a multiplication protocol to calculate the predicted label data. The shared shards corresponding to the predicted label data ;
[0106] for The two parties involved in the calculation use a multiplication agreement to calculate... The numerical comparison protocol is used to obtain elements set These elements are those that do not satisfy the support vector machine constraints;
[0107] Let the learning rate be .from Selected Corresponding data set ,from The In the list choose Corresponding target set Update parameters The List: ;
[0108] Repeat the above steps until the specified number of iterations, then output the shared fragments of the model parameters. .
[0109] In this embodiment of the invention, by using target feature data sharing partitioning and model parameter sharing partitioning for training, it is ensured that the data-stealth learning model to be trained will not directly expose the original data of the participants throughout the process. Furthermore, by using a support vector machine model, it is ensured that the model can effectively capture patterns in the data when processing complex data. By setting a learning rate and updating the model parameters under the guidance of the loss function, the model can dynamically adjust itself to minimize the prediction error, so that the model parameter update can converge effectively and stably, thereby achieving better performance.
[0110] In this embodiment of the invention, optionally, obtaining the shared fragment corresponding to the predicted label data based on the shared fragment of the target feature data and the initial model parameters of the first computational participant includes:
[0111] The shared fragments corresponding to the predicted label data are obtained using a secure multi-party computation protocol based on the target feature data shared fragments and the initial model parameters of the first computation participant. The secure multi-party computation protocol includes at least one of the following: an addition protocol, a multiplication protocol, and a numerical comparison protocol.
[0112] In this embodiment of the invention, in the calculation step of obtaining the shared fragment corresponding to the predicted label data based on the target feature data shared fragment and the initial model parameters of the first computing participant, for secretly shared variables, data anonymity calculation is achieved through a secure multi-party computation protocol.
[0113] For example: the first computation participant is the server, the second computation participant is the client, and when the operand is... Their corresponding secret-sharing fragments are respectively , .
[0114] Both parties can obtain the sum / difference of shared fragments by adding / subtracting the shared fragments of their local operands. That is:
[0115] ;
[0116] If one of the multipliers is public, both parties only need to multiply the shared fragment of the local multiplier by the public multiplier to obtain the shared fragment of the product, that is: ;
[0117] If both multipliers are secretly shared, a multiplication protocol implemented using Beaver triples is required. Please refer to [reference needed]. Figure 5 The specific steps include:
[0118] Random numbers are generated by a semi-honest third party acting as a computational assistant. ,calculate .Will fragments (by fragments Taking the calculation method as an example, before the secret sharing, both parties will agree on the domain of the secret sharing calculation, and It is a number that is uniformly and randomly generated over its domain. Shared separately to the server and client;
[0119] Server-side computing and And send them to the client;
[0120] Client computing and And send them to the server;
[0121] Both sides resumed and ;
[0122] Server-side shared shards for calculating products: ;
[0123] Client-shared shards for calculating products: ;
[0124] Based on the above multiplication protocol, a numerical comparison protocol can be further implemented.
[0125] Please refer to Figure 6 The numerical comparison protocol is as follows: assuming operands Less than a constant .
[0126] The server and client compute locally respectively. and ;
[0127] The server randomly selects two positive integers. and The constraints are satisfied. ; To prevent overflow, the two numbers are set to a large integer and secretly shared, so that both parties receive a shared fragment of the random number. and ;
[0128] Both parties used the above multiplication agreement to calculate and ;
[0129] The server will Send to the client;
[0130] Client recovery as well as , and The comparison result is and The comparison results.
[0131] In this embodiment of the invention, a secure multi-party computation protocol enables computation participants to jointly compute shared fragments of predicted label data without exposing the original data, thereby protecting the privacy of computation participants and preventing the leakage of sensitive information, and improving the security of data analysis.
[0132] Please refer to Figure 7 This invention provides a data anonymity learning method, including:
[0133] Step 71: Obtain the feature data of the target dataset;
[0134] Step 72: Obtain the target shared fragments based on the feature data of the target dataset;
[0135] Step 73: Input the target shared fragment into the data anonymity learning model, and the data anonymity learning model outputs the shared fragment of the label corresponding to the feature data based on the target shared fragment.
[0136] In this embodiment of the invention, the specific steps include:
[0137] The participants in the calculation use a multiplication protocol to calculate And calculate the predicted value, that is, the shared fragment of the label corresponding to the feature data. ;
[0138] if These are the model parameters of the support vector machine. ;if If the parameters are those of a logistic regression model, then both parties involved in the calculation need to use a numerical comparison protocol to determine them. Each element in The interval in which it is located, namely:
[0139] ;
[0140] In some embodiments, a computing participant To obtain the plaintext predicted value for each element in the dataset, the server and client will each create their own prediction shards. Send to each , making Recover the plaintext of the predicted value .
[0141] In this embodiment of the invention, shared fragments are generated for the feature data of the target dataset, so that the data remains encrypted during the flow and processing. The target shared fragments are then input into the data anonymity learning model, which can better adapt to the feature data and output corresponding labeled shared fragments. By generating shared fragments and inputting them into the model, the participants can collaborate effectively while maintaining the privacy of their own data. Multiple participants can share the model's results, ensuring that all participants trust the results.
[0142] Please refer to Figure 8 This invention provides a training apparatus for a data anonymity learning model, comprising:
[0143] The first acquisition module 81 is used to acquire target feature data sharing fragments and target label data sharing fragments. The target feature data sharing fragments and target label data sharing fragments are obtained by extracting the sharing fragments corresponding to the target feature data and target label data from the intersection of the first computing participant sharing fragments and the intersection of the second computing participant sharing fragments. The intersection of the first computing participant sharing fragments and the intersection of the second computing participant sharing fragments are obtained from the secret sharing fragments obtained after performing anonymity and privacy intersection calculation on the associated data of elements with the same unique identifier in the first computing participant dataset and the second computing participant dataset.
[0144] The first processing module 82 is used to input the target feature data shared slice into the data escaping learning model to be trained. The data escaping learning model to be trained obtains the shared slice corresponding to the predicted label data based on the target feature data shared slice and the initial model parameters of the first computing participant. The model parameters of the data escaping learning model are optimized based on the shared slice corresponding to the predicted label data, the target label data shared slice, and the loss function to obtain the trained escaping learning model.
[0145] The training device for the data concealment learning model provided in this embodiment of the invention can achieve Figure 1 The various processes implemented in the method embodiments achieve the same technical effect, and will not be described again here to avoid repetition.
[0146] Please refer to Figure 9 This invention provides a data anonymity learning device, comprising:
[0147] The second acquisition module 91 is used to acquire feature data of the target dataset;
[0148] The second processing module 92 is used to obtain the target shared fragment based on the feature data of the target dataset;
[0149] The third processing module 93 is used to input the target shared fragment into the data anonymity learning model, and the data anonymity learning model outputs the shared fragment of the label corresponding to the feature data according to the target shared fragment.
[0150] The data stealth learning device provided in this embodiment of the invention can achieve Figure 7 The various processes implemented in the method embodiments achieve the same technical effect, and will not be described again here to avoid repetition.
[0151] This invention provides an electronic device 100, see [link to documentation]. Figure 10 As shown, Figure 10 This is a schematic diagram of an electronic device 100 according to an embodiment of the present invention, including a processor 101, a memory 102, and a program or instructions stored in the memory 102 and executable on the processor 101. When the program or instructions are executed by the processor, they implement the training method of any data stealth learning model of the present invention, or implement the steps in any data stealth learning method of the present invention.
[0152] This invention provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they represent the training method of any data-stealth learning model of this invention, or implement various processes of any embodiment of the data-stealth learning method of this invention, and can achieve the same technical effect. To avoid repetition, further details are omitted here.
[0153] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the above-described... Figure 1 or Figure 7 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.
[0154] Computer-readable media include both permanent and non-permanent, removable and non-removable media, which can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0155] It should be noted that the collection, gathering, updating, analysis, processing, use, transmission, and storage of user personal information involved in this disclosed technical solution all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.
[0156] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0157] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a service classification device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0159] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A training method for a data-hidden learning model, characterized in that, The data anonymity learning model is applied to at least one first computational participant and at least one second computational participant, and the training method includes: The first computational participant obtains a shared fragment of target feature data and a shared fragment of target label data. The shared fragments of target feature data and target label data are obtained by extracting the shared fragments corresponding to the target feature data and target label data from the intersection of the first computational participant's shared fragments and the intersection of the second computational participant's shared fragments. The intersection of the first computational participant's shared fragments and the intersection of the second computational participant's shared fragments is obtained from the secret shared fragments obtained after performing anonymity and privacy calculation on the associated data of elements with the same unique identifier in the first computational participant's dataset and the second computational participant's dataset. The first computational participant inputs the target feature data shared slice into the data anonymity learning model to be trained. The data anonymity learning model to be trained obtains the shared slice corresponding to the predicted label data based on the target feature data shared slice and the initial model parameters of the first computational participant. The model parameters of the data anonymity learning model are then optimized based on the shared slice corresponding to the predicted label data, the target label data shared slice, and the loss function to obtain the trained anonymity learning model.
2. The training method for the data anonymity learning model according to claim 1, characterized in that, The data anonymity learning model to be trained obtains the shared slices corresponding to the predicted label data based on the shared slices of the target feature data and the initial model parameters of the first computational participant. Then, it optimizes the model parameters of the data anonymity learning model based on the shared slices corresponding to the predicted label data, the shared slices of the target label data, and the loss function to obtain the trained anonymity learning model, including: The first computational participant trains the data anonymity learning model to be trained using a logistic regression model, specifically including: The first computational participant obtains initial model parameters through the logistic regression model and obtains the model parameter sharing slices corresponding to the initial model parameters; divides the target feature data sharing slices into multiple batches; for each batch of the target feature data sharing slices, a computation process is executed, the computation process including: obtaining the shared slices corresponding to the predicted label data based on the target feature data sharing slices and the model parameter sharing slices; calculating the gradient of the loss function with respect to the model parameters based on the shared slices corresponding to the predicted label data and the target label data sharing slices, performing gradient descent, and updating the model parameters, until the computation process has been executed for each batch of the target feature data sharing slices and a preset number of iterations has been reached, outputting the final model parameter sharing slices, and obtaining the trained stealth learning model based on the final model parameter sharing slices.
3. The training method for the data anonymity learning model according to claim 1, characterized in that, The data anonymity learning model to be trained obtains the shared slices corresponding to the predicted label data based on the shared slices of the target feature data and the initial model parameters of the first computational participant. Then, it optimizes the model parameters of the data anonymity learning model based on the shared slices corresponding to the predicted label data, the shared slices of the target label data, and the loss function to obtain the trained anonymity learning model, including: The first computational participant trains the data anonymity learning model to be trained using a support vector machine model, specifically including: The first computational participant obtains initial model parameters through a support vector machine (SVM) model and obtains a shared model parameter slice corresponding to the initial model parameters. A target batch is obtained by randomly sampling from the shared target feature data slice. The shared target feature data slice containing elements in the target batch that do not satisfy the SVM constraints is extracted, and a computation process is executed. This computation process includes: obtaining a shared slice corresponding to the predicted label data based on the shared target feature data slice and the shared model parameter slice; setting the learning rate of the SVM model; updating the model parameters based on the learning rate, the shared target feature data slice, the shared slice corresponding to the predicted label data, and the shared target label data slice until a preset iteration round is reached; outputting the final shared model parameter slice; and obtaining the trained stealth learning model based on the final shared model parameter slice.
4. The training method for the data anonymity learning model according to claim 1, characterized in that, The step of obtaining the shared fragment corresponding to the predicted label data based on the shared fragment of the target feature data and the initial model parameters of the first calculation participant includes: The shared fragments corresponding to the predicted label data are obtained using a secure multi-party computation protocol based on the target feature data shared fragments and the initial model parameters of the first computation participant. The secure multi-party computation protocol includes at least one of the following: an addition protocol, a multiplication protocol, and a numerical comparison protocol.
5. The training method for the data anonymity learning model according to claim 1, characterized in that, The first computational participant acquires the target feature data sharing fragment and the target label data sharing fragment, including: Obtain a first computational participant dataset, wherein each element in the first computational participant dataset includes: a unique identifier and associated data; The unique identifier of the first computing participant's dataset is processed using a hash function to obtain a first hash value. The first hash value is then multiplied and the associated data is encrypted to obtain a first encrypted dataset. After the first encrypted dataset is randomly rearranged, it is sent to the second computing participant. The system receives a second encrypted dataset sent by the second computing participant and a shared fragment ciphertext of the associated data of the first computing participant. The second encrypted dataset is obtained by the second computing participant processing the unique identifier of the dataset using a hash function to obtain a second hash value, multiplying the second hash value, and then encrypting the associated data. The shared fragment ciphertext of the associated data of the first computing participant is obtained by multiplying the first hash value in the first encrypted dataset, encrypting a first random number, subtracting the encrypted first random number from the associated data of the first encrypted dataset, and finally randomly rearranging the data. The second hash value in the second encrypted dataset is exponentially multiplied to obtain the shared fragmented ciphertext of the data associated with the second computation participants. Extract the shared fragments of the associated data of elements with equal hash values from the shared fragment ciphertext of the first computation participant's associated data and the shared fragment ciphertext of the second computation participant's associated data to obtain the first intersection; Encrypt a second random number, subtract the encrypted second random number from the associated data of the second computing participant in the first intersection to obtain the shared fragment ciphertext of the associated data of the second computing participant in the intersection, and send the shared fragment ciphertext of the associated data of the second computing participant to the second computing participant. The second computing participant decrypts the shared fragment ciphertext of the associated data of the second computing participant and merges the decrypted shared fragment ciphertext of the associated data of the second computing participant with the first random number to obtain the shared fragment of the second computing participant in the intersection. Based on the elements in the first intersection, the shared fragment ciphertext of the first computational participant's associated data is decrypted, and the decrypted shared fragment ciphertext of the first computational participant's associated data is merged with the second random number to obtain the first computational participant's shared fragment of the intersection.
6. The training method for the data anonymity learning model according to claim 5, characterized in that, The first random number and the second random number are uniformly and randomly selected from the plaintext space of the homomorphic encryption framework agreed upon by the first computing participant and the second computing participant.
7. A data anonymity learning method, characterized in that, include: Obtain feature data from the target dataset; The target shared fragments are obtained based on the feature data of the target dataset; The target shared fragment is input into the data anonymity learning model, and the data anonymity learning model outputs the shared fragment of the label corresponding to the feature data based on the target shared fragment.
8. A training device for a data-hidden learning model, characterized in that, include: The first acquisition module is used by the first computing participant to acquire the target feature data sharing fragment and the target label data sharing fragment. The target feature data sharing fragment and the target label data sharing fragment are obtained by extracting the sharing fragments corresponding to the target feature data and the target label data from the intersection of the first computing participant sharing fragment and the intersection of the second computing participant sharing fragment. The intersection of the first computing participant sharing fragment and the intersection of the second computing participant sharing fragment are obtained from the secret sharing fragment obtained after performing anonymity and privacy calculation on the associated data of elements with the same unique identifier in the first computing participant dataset and the second computing participant dataset. The first processing module is used to input the target feature data shared slice into the data anonymity learning model to be trained. The data anonymity learning model to be trained obtains the shared slice corresponding to the predicted label data based on the target feature data shared slice and the initial model parameters of the first computing participant. The model parameters of the data anonymity learning model are optimized based on the shared slice corresponding to the predicted label data, the target label data shared slice, and the loss function to obtain the trained anonymity learning model.
9. A data-anonymized learning device, characterized in that, include: The second acquisition module is used to acquire feature data of the target dataset; The second processing module is used to obtain the target shared fragments based on the feature data of the target dataset. The third processing module is used to input the target shared fragment into the data anonymity learning model, and the data anonymity learning model outputs the shared fragment of the label corresponding to the feature data based on the target shared fragment.
10. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein when the program or instructions are executed by the processor, they implement the training method of the data stealth learning model as described in any one of claims 1 to 6, or the steps in the data stealth learning method as described in claim 7.
11. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the training method of the data stealth learning model as described in any one of claims 1 to 6, or the steps in the data stealth learning method as described in claim 7.
12. A computer program product, characterized in that, It includes computer instructions that, when executed by a processor, implement the training method of the data-stealth learning model as described in any one of claims 1 to 6, or the steps in the data-stealth learning method as described in claim 7.