A privacy protection method and device based on secret sharing and multi-party vertical logistic regression

Abstract

This invention discloses a privacy-preserving method and apparatus for multi-party longitudinal logistic regression based on secret sharing. The method involves a coordinator generating a large number of Beaver triples and distributing them to all participants before the protocol begins. First, model weights are allocated to each participant using secret sharing technology. Then, each participant collaboratively calculates the forward propagation value cross terms. Next, based on their respective datasets, model weight slices, and secret slices of the forward propagation value cross terms, the output value is obtained, and the active party calculates the final forward propagation value. Then, the active and passive parties collaboratively calculate the model gradient cross terms and obtain the secret slices of the model gradient cross terms, and then locally update the model weight slices. Because this invention generates a large number of Beaver triples and distributes them to each participant before the protocol begins, the communication and computational workload can be shifted to before the protocol starts. Furthermore, secret slicing can be performed among multiple participants, thus allowing for expansion to three or more participants.

Description

Technical Field

[0001] This invention relates to the field of information security technology, specifically to a privacy protection method and apparatus for multi-party vertical logistic regression based on secret sharing, which can be used for secure and practical logistic regression analysis. Background Technology

[0002] With increasing focus on data privacy, the question of how to build machine learning (ML) models on different data sources while protecting privacy is gaining traction. Vertical federated learning (VFL) considers the scenario where different stakeholders jointly train an ML model on partitioned features, making it suitable for many real-world cross-enterprise collaborations. Logistic regression (LR) is a classic algorithm in machine learning, and due to its simplicity and effectiveness in many binary classification tasks, privacy-preserving model training or inference based on vertically distributed datasets has attracted widespread attention.

[0003] Existing privacy-preserving solutions for longitudinal logistic regression can be categorized into two types based on how private features are handled. The first type utilizes secure multi-party computation (MPC) techniques, such as homomorphic encryption (HE) and secret sharing (SS), to achieve complete privacy guarantees. Typically, they leverage data outsourcing techniques widely used in database services, outsourcing the dataset to a non-colluding server for ML training or inference. To maintain privacy, all feature values ​​are converted to HE or SS variables during outsourcing, so the server cannot know the original values. However, this approach is not suitable for many real-world datasets. For high-dimensional and sparse datasets, outsourced features become completely dense because non-zero feature indices are also private. This prevents sparse computation, leading to performance risks at high sparsity. The second type follows a split-learning paradigm, which does not outsource the original dataset and therefore handles sparse features well. Typically, each party maintains a low-level model in plaintext, which extracts forward activations using its own private features. The activations of each party are exchanged and fed into the top-level model for prediction. However, the values ​​generated in the low-level model exist in plaintext, leading to data leakage. Such data breaches are prohibited, and even illegal, in real-world applications.

[0004] In 2022, Fu, Fangcheng, et al. proposed BlindFL, a secure and practical privacy-preserving scheme for vertical logistic regression based on HE and SS (the work "BlindFL: Vertical Federated Machine Learning without Peeking into Your Data." was published in the Proceedings of the 2022 International Conference on Management of Data). BlindFL keeps the private dataset within each party without outsourcing and unifies the functionality through an important component called the "Source Layer." By doing so, BlindFL can support sparse input features while achieving security guarantees. Figure 1 As shown, the BlindFL implementation can be decomposed into two parts: forward propagation and backward propagation. During forward propagation, the joint source layer serves as the basic building block, combining the features of passive agent A and active agent B (which possesses the true label values). Its output is represented as Z, assuming that A and B possess datasets X and X respectively. A X B That is, Z = X A W A +X B W B The non-federated submodule of B, namely the top model, acts as the classifier / predictor. During backpropagation, B calculates the loss function using the final predicted values ​​and the true labels, and then backpropagates along the top model to obtain the partial derivative of the loss function with respect to Z. Furthermore, the model gradient can be calculated using the chain rule. Finally, the model weights of the source layer are updated using the model gradient.

[0005] Specifically, to prevent the leakage of the datasets held by A and B, as well as B's label values, BlindFL prohibits either party from possessing their own or the other party's underlying model, W. A and W B During forward propagation, the passive party A is prohibited from revealing the forward propagation value Z. During backward propagation, A and B cannot reveal the gradient of the other party's model, and A cannot reveal its own model gradient either. To achieve these restrictions, BlindFL uses SS to store W... A and W B The secret is shared with A and B, but A and B only have knowledge of W. A and W BTo calculate the forward propagation value Z, A and B need to homomorphically encrypt one of their fragments and send it to the other. They then perform a homomorphic scalar multiplication between the obtained ciphertext fragment and their own plaintext fragment, transforming the result into a secret-sharing form. This process requires multiple rounds of communication between A and B, with B ultimately only obtaining the forward propagation value Z. The backward propagation process is similar; after multiple rounds of communication, A and B can obtain the value used to update W. A and W B Piecewise model gradient The fragments.

[0006] However, the BlindFL scheme has insufficient scalability. When extended to multiple participants, the BlindFL scheme needs to be redesigned. In addition, the scheme requires too many communication rounds, especially when extended to multiple parties, requiring pairwise communication, which results in a huge time overhead. Summary of the Invention

[0007] This invention addresses the technical problems of existing longitudinal logistic regression privacy protection schemes, such as difficulty in scaling to multiple parties and the large number of communication rounds. It proposes a multi-party longitudinal logistic regression privacy protection method and apparatus based on secret sharing. This scheme can be used for secure multi-party longitudinal logistic regression and achieves fewer communication rounds. Because the method of this invention does not outsource the original data, it is suitable for high-dimensional and sparse datasets. Furthermore, each party cannot access any of the underlying model, thus achieving the same security guarantees as the aforementioned schemes.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] The first aspect provides a multi-party longitudinal logistic regression privacy protection method based on secret sharing, applicable to multiple participants, including active and passive parties, each possessing their own dataset. A facilitator is responsible for generating a large number of Beaver triples and distributing them to each participant before the protocol begins. The multi-party longitudinal logistic regression privacy protection method includes:

[0010] The model weights are allocated to each participant using a secret sharing technique.

[0011] Each participant collaboratively calculates the forward propagation value cross term based on the assigned model weight slice and ternary component slice, and obtains the secret slice of the forward propagation value cross term, wherein the forward propagation value cross term is a part of the calculation of the forward propagation value;

[0012] Each participant obtains the output value based on its own dataset, model weight shards, and secret shards of the forward propagation value interaction terms. The passive party sends the obtained output value to the active party, which obtains the forward propagation value based on the passive party's output value and its own output value. The forward propagation value serves as the input to the top-level model and can obtain the model's final prediction value.

[0013] The active and passive parties collaboratively compute the model gradient cross term and obtain the secret slice of the model gradient cross term. Each participant locally updates the model weight slice based on the secret slice of the model gradient cross term. The model gradient cross term is a part of the model gradient that needs to be collaboratively computed by the active and passive parties.

[0014] In one implementation, model weights are allocated to each participant using a secret-sharing technique, including:

[0015] The model weights are distributed to each participant by W = U + v. The model weight slices U and v are held by different participants, where W is the model weight.

[0016] In one implementation, each participant collaboratively calculates the forward propagation value cross term based on the assigned model weight slice and ternary component slice, and obtains the secret slice of the forward propagation value cross term, including:

[0017] Each participant reconstructs the parameters used to calculate the cross term based on the assigned model weights and ternary components, and calculates the secret component of the forward propagation value cross term based on the parameters used to calculate the cross term and the ternary components.

[0018] In one implementation, each participant reconstructs the parameters for calculating the cross term based on the assigned model weight slices and ternary slices, and calculates the secret slice of the forward propagation value cross term based on the parameters for calculating the cross term and the ternary slices, including:

[0019] Each participant initializes model weight sharding and dataset sharding based on the assigned model weights and the datasets they possess;

[0020] Each participant calculates the shards for calculating the parameters of the cross term based on the initialized model weight shards, dataset shards, and triplet shards, and calculates the parameters for calculating the cross term based on the shards for calculating the parameters of the cross term.

[0021] Each participant calculates the secret slice of the forward propagation value cross term based on the parameters used to calculate the cross term and the ternary component slice.

[0022] In one implementation, each participant obtains the output value based on its own dataset, model weight sharding, and cross-term sharding, including:

[0023] Each participant multiplies its own dataset with the model weight slices and then adds the result to the slices of the interaction terms to obtain the output value.

[0024] In one implementation, the active party obtains the forward propagation value based on the passive party's output value and its own output value, including:

[0025] The active party adds the received output value from the passive party to its own output value to obtain the forward propagation value.

[0026] In one implementation, the active and passive parties collaboratively compute model gradient cross terms and obtain secret slices of these cross terms, including:

[0027] The active and passive parties initialize error value sharding and dataset sharding based on the error value between the predicted label and the true label and the dataset they have.

[0028] The active and passive sides calculate the partitions used to calculate the model gradient cross terms based on the initialized partitions and ternary partitions, and use the partitions used to calculate the model gradient cross terms as parameters for calculating the model gradient.

[0029] The active and passive parties calculate the secret slices of the model gradient cross terms based on the parameters used to calculate the model gradient and the ternary component slices.

[0030] Based on the same inventive concept, a second aspect of the present invention provides a multi-party longitudinal logistic regression privacy protection device based on secret sharing, applied to multiple participating parties, including active and passive parties, each party possessing its own dataset. An facilitator is responsible for generating a large number of Beaver triples and distributing them to each participating party before the protocol begins. The multi-party longitudinal logistic regression privacy protection device includes:

[0031] The model weight sharing module is used to distribute model weights to each participant using secret sharing technology;

[0032] The forward propagation value cross term secret sharding calculation module is used by each participant to collaboratively calculate the forward propagation value cross term according to the allocated model weight sharding and triplet sharding, and obtain the secret shard of the forward propagation value cross term, wherein the forward propagation value cross term is a part of the calculation of the forward propagation value;

[0033] The forward propagation value calculation module is used by each participant to obtain the output value based on their own dataset, model weight shards, and secret shards of the forward propagation value intersection. The passive party sends the obtained output value to the active party, and the active party obtains the forward propagation value based on the passive party's output value and its own output value. The forward propagation value serves as the input to the top-level model and can obtain the model's final prediction value.

[0034] The model weight shard update module is used by the active and passive parties to collaboratively compute the model gradient cross term and obtain the secret shard of the model gradient cross term. Each participating party updates the model weight shard locally based on the secret shard of the model gradient cross term. The model gradient cross term is a part of the model gradient that needs to be collaboratively computed by the active and passive parties.

[0035] Based on the same inventive concept, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-party vertical logical regression privacy protection method based on secret sharing as described in the first aspect.

[0036] Based on the same inventive concept, a fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the multi-party vertical logical regression privacy protection method based on secret sharing as described in the first aspect.

[0037] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0038] This invention proposes a privacy-preserving method for multi-party longitudinal logistic regression based on secret sharing. Before the protocol begins, a coordinator generates a large number of Beaver triples and distributes them to all participants. Secret sharing is used to allocate model weights to each participant. During forward propagation, the passive party collaboratively calculates the forward propagation value cross terms based on the allocated model weight slices and triple slices, obtaining secret slices of these cross terms. Each participant then obtains the output value based on their own dataset, model weight slices, and cross term slices. The active party ultimately obtains the forward propagation value. During backpropagation, the active and passive parties collaboratively calculate the model gradient cross terms and obtain secret slices of these cross terms. Each participant locally updates its model weight slices based on these secret slices. Because this method generates a large number of Beaver triples and distributes them to all participants before the protocol begins, the communication and computational workload is shifted to before the protocol starts, significantly reducing the computational and communication workload during collaborative computation. Furthermore, secret slicing can be performed among multiple participants, thus extending to three or more participants. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart of the BlindFL longitudinal logistic regression privacy protection scheme;

[0041] Figure 2 A flowchart of a multi-party longitudinal logistic regression privacy protection method based on secret sharing provided in an embodiment of the present invention;

[0042] Figure 3 This is a structural diagram of a multi-party vertical logistic regression privacy protection device based on secret sharing, provided in an embodiment of the present invention. Detailed Implementation

[0043] The inventors of this application discovered through extensive research and practice that Figure 1 The vertical logistic regression privacy protection scheme described in the article mainly has the following two problems:

[0044] (1) It is not easy to extend to multiple participants;

[0045] (2) There are many communication rounds.

[0046] This invention aims to solve the two problems mentioned above. The proposed novel scheme can be used for secure multi-party longitudinal logistic regression and achieves fewer communication rounds. Since the method of this invention does not outsource the original data, it is suitable for high-dimensional and sparse datasets. At the same time, each party cannot access any of the underlying model, thus achieving the same security guarantee as the schemes mentioned above.

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 embodiments of the present invention, not all embodiments. 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.

[0048] Example 1

[0049] This invention discloses a privacy-preserving method for multi-party longitudinal logistic regression based on secret sharing, applicable to multiple participants, including active and passive parties, each possessing their own dataset. A facilitator is responsible for generating a large number of Beaver triples and distributing them to each participant before the protocol begins. (See also...) Figure 2 The method includes:

[0050] S1: Distribute model weights to each participant using secret sharing technology;

[0051] S2: Each participant collaboratively calculates the forward propagation value cross term based on the assigned model weight slice and ternary component slice, and obtains the secret slice of the forward propagation value cross term, wherein the forward propagation value cross term is a part of the calculation of the forward propagation value;

[0052] S3: Each participant obtains the output value based on its own dataset, model weight shards, and secret shards of the forward propagation value intersection. The passive party sends the obtained output value to the active party. The active party obtains the forward propagation value based on the passive party's output value and its own output value. The forward propagation value serves as the input to the top-level model and can obtain the model's final prediction value.

[0053] S4: The active and passive parties jointly compute the model gradient cross term and obtain the secret slice of the model gradient cross term. Each participant locally updates the model weight slice based on the secret slice of the model gradient cross term. The model gradient cross term is a part of the model gradient that needs to be computed jointly by the active and passive parties.

[0054] Specifically, step S1 is the initialization step, in which the participants share model weights through secret sharing technology. S2 and S3 belong to the forward propagation process. In S2, the participants collaboratively calculate the forward propagation value cross term and obtain the secret slice of the forward propagation value cross term. In S3, each participant obtains the output value based on their own dataset, model weight slice, and secret slice of the forward propagation value cross term, and the active party calculates the forward propagation value. S4 belongs to the back propagation process, in which the active party and the passive party collaboratively calculate the model gradient cross term and obtain the secret slice of the model gradient cross term. Each participant locally updates the model weight slice based on the secret slice of the model gradient cross term.

[0055] Logistic regression is a classic algorithm in machine learning. Inspired by a research paper, this invention proposes a privacy-preserving method for multi-party longitudinal logistic regression based on secret sharing. This method can be easily extended to three or more parties, achieving a low number of communication rounds while ensuring security. The Beaver triple is a crucial basic component of this scheme. The algorithm flow of the Beaver triple is briefly described first, followed by a principle description of the secret-sharing-based multi-party longitudinal logistic regression privacy-preserving algorithm using a three-party example.

[0056] Use [*] to denote the state after the secret has been shared, such as [a] indicating that secret 'a' has been shared with the participants through the secret sharing function. For a two-party secret sharing, record two fragment values ​​[a]. o and Right now

[0057] The Beaver triplet algorithm is described as follows:

[0058] Before the protocol begins, a random triplet is generated in advance: [a], [b], [c]. c satisfies c = a·b. Assuming that secrets x and y have already been shared, each participant now needs to compute [x·y], which is a secret slice of the product of x and y.

[0059] (1) All participants calculate and publish [α], [α] = [x] - [a].

[0060] (2) All participants calculate and publish [β], [β] = [y] - [b].

[0061] (3) All participants calculate [z], where [z] o = [c] + α·[b] + β·[a],

[0062] Since [a], [b], and [c] are pre-generated and allocated triples, during multiplication, participants only need to locally calculate [α] = [x] - [a] and [β] = [y] - [b], and publicly disclose the results [α] and [β]. All participants can independently calculate α and β through a secret reconstruction function. Beaver triples are consumable; each multiplication operation consumes one Beaver triple. By using pre-computed Beaver triples, communication and computational overhead can be shifted to before the protocol begins.

[0063] It should be noted that [x·y] represents the secret slice of x·y. For simplicity, the result of the calculation [x]-[a] is denoted as [α], [β], and [z]. The triples held by each participant are the slices of a, b, and c ([a], [b], and [c]), because they have the relationship c = a·b, hence the term triples. [z] represents a general slice of secret sharing. o , The subscript indicates the slice held by each participant. For example, if z is secretly shared between two parties and z = 6, then the slice [z] of z held by participant 1 is denoted as [z]. o =1, the fragment [z] of z's secret shared by participant 2 is denoted as In addition, [a] o and The difference is that these two shards are controlled by different participants.

[0064] In practice, the privacy protection method based on secret sharing and multi-party vertical logistic regression is described as follows:

[0065] Suppose there are three participants involved in training the model, denoted as P1, P2, and P3, where P1 and P2 are passive participants, and P3 is the active participant. Their datasets are X1, X2, and X3, respectively. An facilitator, P0, is responsible for generating a large number of Beaver triples and distributing them to each participant before the protocol begins. In plaintext mode, the underlying model weights for each participant are denoted as W1, W2, and W3, respectively.

[0066] 1. Initialization

[0067] To ensure that no party possesses any of the underlying model's underlying structure—that is, no party owns the plaintext model weights W1, W2, and W3—the model weights are first distributed among the participants using SS (Secret Sharing) technology. In practice, the model weights can be secretly shared using W = U + v, where the weight fragments U and v are held by different participants, so none of them know the value of W. To achieve this, during initialization, P... i Initialize U i and V i-1 , i = {1, 2, 3}.

[0068] After initialization, the data held by each participant is as follows:

[0069] Participants P1: U1, V3, X1;

[0070] Participants P2: U2, V1, X2;

[0071] Participants P3: U3, V2, X3;

[0072] 2. Forward propagation

[0073] Objective: The active party P3 obtains the value of Z. The value of Z is the forward activation value of the underlying model.

[0074] Z=X1W1+X2W2+X3W3=X1(U1+V1)+X2(U2+V2)+X3(U3+V3)

[0075] =X1U1+X1V1+X2U2+X2V2+X3U3+X3V3

[0076] Among them, X1U1, X2U2, and X3U3 can be calculated locally by P1, P2, and P3, while X1V1, X2V2, and X3V3 are cross terms and require knowledge of X. i and V i Both parties use Beaver triples for collaborative computation, and the final result is that both parties can obtain secret fragments of the cross terms.

[0077] Taking the cross term X1V1 as an example, X1 is owned by P1, and P1 initializes the fragment of X1 [X1] = X1, while P2 initializes the fragment of X1 [X1] = 0. V1 is owned by P2, and P2 initializes the fragment of V1 [V1] = V1, while P1 initializes the fragment of V1 [V1] = 0. P0 has already generated and assigned random triples to P1 and P2 before the protocol started, denoted as [a], [b], and [c]. P1 and P2 respectively calculate and publish [α] = [X1] - [a] and [β] = [V1] - [b]. After P1 and P2 reconstruct α and β, they can locally calculate [X1V1]. Similarly, the fragments of other cross terms can be calculated.

[0078] The specific calculation steps of the above process (participants P1 and P2 collaboratively calculate the forward propagation value cross term based on the allocated model weights and the ternary component, and obtain the secret slice of the forward propagation value cross term) are as follows:

[0079] (1) P1 initializes [X1]o = X1, P2 initialization [V1] o =V1. ( (These are model weight sharding and dataset sharding, respectively)

[0080] (2) Calculate [α] using P1 o =[X1] o -[a] o , And publicly available, P2 calculation And made public.

[0081] (3) P1 and P2 reconstruct the (α and β are parameters used to calculate the cross term)

[0082] (4) P1 local calculation [X1V1] o =[c] o +α·[b] o +β·[a] o P2 local computing This is a secret slice of the forward propagation value cross term.

[0083] The proof of correctness is as follows:

[0084]

[0085] After the above local calculations, the data held by each participant is as follows:

[0086] Participants P1: U1, V3, X1, X1U1, [X1V1] o ,

[0087] Participants P2: U2, V1, X2, X2U2 [X2V2] o ;

[0088] Participants P3: U3, V2, X3, X3U3 [X3V3] o ;

[0089] Each participating party P i Calculate separately The calculation result is recorded as Z. i Then, P1 and P2 send Z1 and Z2 to P3 respectively. P3 calculates Z locally: Z = Z1 + Z2 + Z3 = X1W1 + X2W2 + X3W3. Using Z as input to the top-level model, the final predicted value of the model can be calculated. The top-level model is unique to P3 and does not require collaborative calculation by all participating parties.

[0090] The specific algorithm for the above process is as follows:

[0091] (1) P1 local calculation P2 local calculation P3 local computing

[0092] (2) P1 sends Z1 to P3, and P2 sends Z2 to P3.

[0093] (3) P3 calculates Z = Z1 + Z2 + Z3.

[0094] 3. Backpropagation

[0095] Objective: Update U1 and V3 in P1, update U2 and V1 in P2, and update U3 and V2 in P3.

[0096] For participant P i The model gradient in plaintext form can be obtained through Calculation. Because P3 obtained the value of Z during forward propagation, P3 can calculate it locally. in This represents the partial derivative of the loss function with respect to Z. Since... This is unique to P3. Similar to the method for calculating cross terms during forward propagation, P1 and P3 need to collaboratively calculate cross terms. To obtain the value and its secret fragment, P2 and P3 need to collaboratively compute the cross term. The value is obtained and its secret slice is acquired. Furthermore, since P2 is the one who holds the weight slice V1 of the P1 model, P3 needs to... Send to P2. Error value sharding and dataset sharding are equivalent to this part. and X i The partitioning, the error value and equivalence.

[0097] The specific algorithm for the above process is as follows:

[0098] (1) P3 local calculation

[0099] (2) Calculate P1 and P3 respectively. and and Secret slices of the model gradient cross terms computed for P1 and P3.

[0100] (3) Calculate P2 and P3 respectively. and and Secret slices of the model gradient cross terms computed for P2 and P3.

[0101] (4) P3 will Send to P2.

[0102] After the aforementioned local calculations and one round of communication, the data held by each participant during backpropagation is as follows:

[0103] Participants P1: U1, V3, X1

[0104] Participants P2: U2, V1, X2

[0105] Participants P3: U3, V2, X3

[0106] Assuming a learning rate of μ, each party locally updates the model weight slices:

[0107] V′3 = V3; U′1 and V′3 are the model weight slices updated by P1;

[0108] U′2 and V′1 are the model weight slices updated by P2;

[0109] U′3 and V′2 are the model weight slices after P3 update.

[0110] Taking W1 as an example, in plaintext form... The correctness of the above update result, W1, is proven as follows:

[0111]

[0112] Example 2

[0113] Based on the same inventive concept, this embodiment discloses a multi-party longitudinal logistic regression privacy protection device based on secret sharing, applied to multiple participating parties, including active and passive parties. Each participating party possesses its own dataset. An facilitator is responsible for generating a large number of Beaver triples before the protocol begins and distributing them to each participating party. Please refer to [link to relevant documentation]. Figure 3 The multi-party vertical logistic regression privacy protection device includes:

[0114] The model weight sharing module 301 is used to distribute model weights to each participant using secret sharing technology;

[0115] The forward propagation value cross term secret sharding calculation module 302 is used by each participant to collaboratively calculate the forward propagation value cross term according to the allocated model weight sharding and ternary group sharding, and obtain the secret shard of the forward propagation value cross term, wherein the forward propagation value cross term is a part of the calculation of the forward propagation value;

[0116] The forward propagation value calculation module 303 is used by each participant to obtain the output value based on their own dataset, model weight shards, and secret shards of the forward propagation value intersection. The passive party sends the obtained output value to the active party, and the active party obtains the forward propagation value based on the passive party's output value and its own output value. The forward propagation value serves as the input to the top-level model and can obtain the model's final prediction value.

[0117] The model weight shard update module 304 is used by the active party and the passive party to collaboratively compute the model gradient cross term and obtain the secret shard of the model gradient cross term. Each participating party updates the model weight shard locally based on the secret shard of the model gradient cross term. The model gradient cross term is a part of the model gradient that needs to be collaboratively computed by the active party and the passive party.

[0118] Since the apparatus described in Embodiment 2 of this invention is the same apparatus used to implement the multi-party vertical logical regression privacy protection method based on secret sharing in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this apparatus based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All apparatuses used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0119] Example 3

[0120] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.

[0121] Since the computer-readable storage medium described in Embodiment 3 of this invention is the same computer-readable storage medium used in implementing the multi-party vertical logical regression privacy protection method based on secret sharing in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer-readable storage medium based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer-readable storage media used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0122] Example 4

[0123] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0124] Since the computer device described in Embodiment 4 of this invention is the same computer device used to implement the multi-party vertical logistic regression privacy protection method based on secret sharing in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0125] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A privacy-preserving method based on secret sharing and multi-party vertical logistic regression, characterized in that, This method applies to multiple participants, including active and passive parties, each possessing their own dataset. An facilitator is responsible for generating a large number of Beaver triples and distributing them to each participant before the protocol begins. The multi-party longitudinal logistic regression privacy protection method includes: The model weights are allocated to each participant using a secret sharing technique. Each participant collaboratively calculates the forward propagation value cross term based on the assigned model weight slice and ternary component slice, and obtains the secret slice of the forward propagation value cross term, wherein the forward propagation value cross term is a part of the calculation of the forward propagation value; Each participant obtains the output value based on its own dataset, model weight shards, and secret shards of the forward propagation value interaction terms. The passive party sends the obtained output value to the active party, which obtains the forward propagation value based on the passive party's output value and its own output value. The forward propagation value serves as the input to the top-level model and can obtain the model's final prediction value. The active and passive parties collaboratively compute the model gradient cross term and obtain the secret slice of the model gradient cross term. Each participant locally updates the model weight slice based on the secret slice of the model gradient cross term. The model gradient cross term is a part of the model gradient that needs to be collaboratively computed by the active and passive parties.

2. The privacy protection method for multi-party vertical logistic regression based on secret sharing as described in claim 1, characterized in that, The model weights are allocated to each participant using a secret-sharing technique, including: The model weights are distributed to each participant using W = U + V. The model weight slices U and V are held by different participants, where W is the model weight.

3. The privacy protection method for multi-party vertical logistic regression based on secret sharing as described in claim 1, characterized in that, Each participant collaboratively calculates the forward propagation value cross term based on the assigned model weight slices and ternary component slices, and obtains the secret slice of the forward propagation value cross term, including: Each participant reconstructs the parameters used to calculate the cross term based on the assigned model weights and ternary components, and calculates the secret component of the forward propagation value cross term based on the parameters used to calculate the cross term and the ternary components.

4. The privacy protection method for multi-party vertical logistic regression based on secret sharing as described in claim 3, characterized in that, Each participant reconstructs the parameters for calculating the cross term based on the assigned model weight slices and triplet slices, and calculates the secret slice of the forward propagation value cross term based on the parameters for calculating the cross term and the triplet slices, including: Each participant initializes model weight sharding and dataset sharding based on the assigned model weights and the datasets they possess; Each participant calculates the shards for calculating the parameters of the cross term based on the initialized model weight shards, dataset shards, and triplet shards, and calculates the parameters for calculating the cross term based on the shards for calculating the parameters of the cross term. Each participant calculates the secret slice of the forward propagation value cross term based on the parameters used to calculate the cross term and the ternary component slice.

5. The privacy protection method for multi-party vertical logistic regression based on secret sharing as described in claim 1, characterized in that, Each participant obtains the output value based on its own dataset, model weight sharding, and interaction term sharding, including: Each participant multiplies its own dataset with the model weight slices and then adds the result to the slices of the interaction terms to obtain the output value.

6. The privacy protection method for multi-party vertical logistic regression based on secret sharing as described in claim 1, characterized in that, The active party obtains forward propagation values ​​based on the passive party's output value and its own output value, including: The active party adds the received output value from the passive party to its own output value to obtain the forward propagation value.

7. The privacy protection method for multi-party vertical logistic regression based on secret sharing as described in claim 1, characterized in that, The active and passive parties collaboratively compute the model gradient cross term and obtain the secret slices of the model gradient cross term, including: The active and passive parties initialize error value sharding and dataset sharding based on the error value between the predicted label and the true label and the dataset they have. The active and passive sides calculate the partitions used to calculate the model gradient cross terms based on the initialized partitions and ternary partitions, and use the partitions used to calculate the model gradient cross terms as parameters for calculating the model gradient. The active and passive parties calculate the secret slices of the model gradient cross terms based on the parameters used to calculate the model gradient and the ternary component slices.

8. A privacy protection device based on secret sharing and multi-party vertical logistic regression, characterized in that, Applied to multiple parties, including active and passive parties, each party possesses its own dataset. An facilitator is responsible for generating a large number of Beaver triples and distributing them to each party before the protocol begins. The multi-party longitudinal logistic regression privacy protection device includes: The model weight sharing module is used to distribute model weights to each participant using secret sharing technology; The forward propagation value cross term secret sharding calculation module is used by each participant to collaboratively calculate the forward propagation value cross term according to the allocated model weight sharding and triplet sharding, and obtain the secret shard of the forward propagation value cross term, wherein the forward propagation value cross term is a part of the calculation of the forward propagation value; The forward propagation value calculation module is used by each participant to obtain the output value based on their own dataset, model weight shards, and secret shards of the forward propagation value intersection. The passive party sends the obtained output value to the active party, and the active party obtains the forward propagation value based on the passive party's output value and its own output value. The forward propagation value serves as the input to the top-level model and can obtain the model's final prediction value. The model weight shard update module is used by the active and passive parties to collaboratively compute the model gradient cross term and obtain the secret shard of the model gradient cross term. Each participating party updates the model weight shard locally based on the secret shard of the model gradient cross term. The model gradient cross term is a part of the model gradient that needs to be collaboratively computed by the active and passive parties.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the multi-party longitudinal logistic regression privacy protection method based on secret sharing as described in any one of claims 1 to 7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multi-party vertical logistic regression privacy protection method based on secret sharing as described in any one of claims 1 to 7.

Images