Business model training system and method based on multi-party secure computation

Abstract

The embodiment of the specification provides a business model training system and method based on multi-party secure calculation, wherein the business model training system based on multi-party secure calculation comprises: a feature providing end configured to process local business features based on a local ciphertext model, obtain a ciphertext prediction result, and send the ciphertext prediction result to a label providing end; the label providing end is configured to receive the ciphertext prediction result, determine intermediate features according to the ciphertext prediction result, construct sample features based on the intermediate features and local associated features, and group a business sample set according to the sample features and local sample labels, train a business model by using the business sample set, and obtain a target business model satisfying a training stop condition.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of artificial intelligence technology, and in particular to a business model training system and method based on multi-party secure computation. Background Technology

[0002] With the development of internet technology, data has gradually become a core asset for various businesses. Especially in today's rapidly developing online business environment, data is indispensable for credit granting, risk control, recommendation systems, and marketing. The quality and quantity of data are the two most important factors affecting the performance of machine learning models. Typically, a machine learning system includes two subsystems: offline training and online prediction. Both stages require a large amount of arithmetic computation. Therefore, when a business possesses limited, low-quality, or insufficient data, multi-party secure computation is often used to improve the prediction accuracy of machine learning models. However, the multiple parties involved in the computation need to be constantly online and synchronously performing calculations, resulting in massive communication volumes. Furthermore, the security of data encryption is vulnerable to breaches. Therefore, an effective solution is urgently needed to address these issues. Summary of the Invention

[0003] In view of this, embodiments of this specification provide a business model training system based on multi-party secure computation. One or more embodiments of this specification also relate to a business processing system based on multi-party secure computation, a business model training method based on multi-party secure computation, a business processing method based on multi-party secure computation, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, a business model training system based on multi-party secure computation is provided, comprising:

[0005] The feature provider is configured to process local business features based on a local ciphertext model, obtain ciphertext prediction results, and send them to the tag provider.

[0006] The tag provider is configured to receive the ciphertext prediction result, determine intermediate features based on the ciphertext prediction result, construct sample features based on the intermediate features and local association features, form a business sample set based on the sample features and local sample tags, and train the business model using the business sample set until a target business model that meets the training stopping condition is obtained.

[0007] According to a second aspect of the embodiments of this specification, a business processing system based on multi-party secure computation is provided, comprising:

[0008] The tag provider is configured to acquire business prediction information, create a business prediction request based on the business prediction information, and send it to the feature provider.

[0009] The feature provider is configured to read the target business features in response to the business prediction request, process the target business features using a local encrypted model, obtain the target encrypted prediction result, and send it to the tag provider.

[0010] The tag provider is also configured to receive the target ciphertext prediction result, determine the target intermediate features based on the target ciphertext prediction result, construct target features based on the target intermediate features and target local association features, and input the target features into the target business model in the above system for processing to obtain business prediction results.

[0011] According to a third aspect of the embodiments of this specification, a method for training a business model based on multi-party secure computation is provided, comprising:

[0012] The local business characteristics are processed based on the local ciphertext model to obtain ciphertext prediction results;

[0013] Intermediate features are determined based on the ciphertext prediction results. Sample features are constructed based on the intermediate features and local association features. A business sample set is formed based on the sample features and local sample labels. The business model is trained using the business sample set until a target business model that meets the training stopping condition is obtained.

[0014] According to a fourth aspect of the embodiments of this specification, a business processing method based on multi-party secure computation is provided, including:

[0015] Obtain business forecast information and create a business forecast request based on the business forecast information;

[0016] In response to the business prediction request, the target business features are read, and the target business features are processed using a local encrypted model to obtain the target encrypted prediction result.

[0017] Based on the target ciphertext prediction result, target intermediate features are determined. Target features are constructed based on the target intermediate features and target local association features. The target features are then input into the target business model in the above method for processing to obtain business prediction results.

[0018] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0019] Memory and processor;

[0020] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement any of the steps of the above-mentioned business model training method based on multi-party secure computation or business processing method based on multi-party secure computation.

[0021] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the above-described business model training method based on multi-party secure computation or business processing method based on multi-party secure computation.

[0022] According to a seventh aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described business model training method based on multi-party secure computation or business processing method based on multi-party secure computation.

[0023] The business model training system based on multi-party secure computation provided in this specification aims to complete model training while reducing communication volume. The feature provider can process local business features based on a local encrypted model, obtain encrypted prediction results, and send them to the tag provider. The tag provider can then determine intermediate features based on the encrypted prediction results, enabling the tag provider to train the model using these intermediate features without disclosing data to the tag provider. Furthermore, the tag provider can construct sample features based on the intermediate features and their associated local features. Then, it combines local sample labels with these sample features to form a business sample set. Using this business sample set to train the business model yields a target business model that meets the business scenario requirements. This process, where the feature provider uses a local encrypted model and the tag provider combines encrypted prediction results from other ends to construct the sample set, completes business model training with minimal communication, effectively reducing communication volume while ensuring data security at all ends. This ensures that the trained target business model meets the needs of the business scenario. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of a business model training system based on multi-party secure computation, provided in one embodiment of this specification.

[0025] Figure 2 This is a schematic diagram of model training and model prediction in a business model training system based on multi-party secure computation, provided in one embodiment of this specification.

[0026] Figure 3 This is a schematic diagram of the structure of a business processing system based on multi-party secure computation, provided in one embodiment of this specification.

[0027] Figure 4This is a schematic diagram of a business processing system based on multi-party secure computation provided in one embodiment of this specification;

[0028] Figure 5 This is a flowchart illustrating a business model training method based on multi-party secure computation, provided in one embodiment of this specification.

[0029] Figure 6 This is a flowchart illustrating a business processing method based on multi-party secure computation, provided in one embodiment of this specification.

[0030] Figure 7 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0031] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0032] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0033] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0034] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0035] Secure multi-party computation: Also known as secure multi-party computation, a typical application of secure multi-party computation is the joint statistical analysis and machine learning of privacy-preserving multi-party data. Secure multi-party computation allows participating parties to compute statistical and machine learning results based on joint data without exposing their individual raw data. The function in secure multi-party computation is either a statistical operation function or a machine learning algorithm.

[0036] Homomorphic encryption is a cryptographic technique based on the computational complexity theory of mathematical problems. Processing homomorphically encrypted data to obtain an output, and then decrypting that output, yields the same result as processing the unencrypted original data using the same method. In other words, processing the ciphertext directly yields the same result as processing the plaintext and then encrypting the result.

[0037] Stacking refers to training a model to combine the predictions of several other models. It involves training some sub-models using available data, and then training a top-level model to use all the predictions from other algorithms as additional input for the final prediction. In other words, multiple sub-models are first trained using data and multiple algorithms. After training, the data to be processed is input into these trained sub-models. The outputs of these trained models, along with the actual results corresponding to the data to be processed, form a new dataset. This new dataset is then used to train another model, ultimately creating a stacked ensemble training model.

[0038] This specification provides a business model training system based on multi-party secure computation. One or more embodiments of this specification also relate to a business processing system based on multi-party secure computation, a business model training method based on multi-party secure computation, a business processing method based on multi-party secure computation, a computing device, a computer-readable storage medium, and a computer program, which will be described in detail in the following embodiments.

[0039] In practical applications, machine learning systems consist of two subsystems: offline training and online prediction. Both require a large number of arithmetic and logical operations, such as matrix multiplication, ReLU, sigmoid, and softmax. Especially in scenarios involving multi-party joint training of models, these operations can be implemented using corresponding MPC (Mobility Personal Computer) protocols, but these involve significant interactions. In other words, all parties involved in the computation need to be constantly online and performing synchronous operations, resulting in substantial communication overhead. Furthermore, the high frequency of communication reduces data security. Therefore, an effective solution is urgently needed to address these issues.

[0040] The business model training system based on multi-party secure computation provided in this specification aims to complete model training while reducing communication volume. The feature provider can process local business features based on a local encrypted model, obtain encrypted prediction results, and send them to the tag provider. The tag provider can then determine intermediate features based on the encrypted prediction results, enabling the tag provider to train the model using these intermediate features without disclosing data to the tag provider. Furthermore, the tag provider can construct sample features based on the intermediate features and their associated local features. Then, it combines local sample labels with these sample features to form a business sample set. Using this business sample set to train the business model yields a target business model that meets the business scenario requirements. This process, where the feature provider uses a local encrypted model and the tag provider combines encrypted prediction results from other ends to construct the sample set, completes business model training with minimal communication, effectively reducing communication volume while ensuring data security at all ends. This ensures that the trained target business model meets the needs of the business scenario.

[0041] Figure 1 This is a schematic diagram of a business model training system based on multi-party secure computation according to one embodiment of this specification. The business model training system 100 based on multi-party secure computation includes a feature providing end 110 and a label providing end 120. The system includes:

[0042] Feature provider 110 is configured to process local business features based on a local ciphertext model, obtain ciphertext prediction results, and send them to tag provider;

[0043] The tag providing end 120 is configured to receive the ciphertext prediction result, determine intermediate features based on the ciphertext prediction result, construct sample features based on the intermediate features and local association features, form a business sample set based on the sample features and local sample tags, and train the business model using the business sample set until a target business model that meets the training stopping condition is obtained.

[0044] Specifically, the feature provider refers to the business entity that holds relatively simple business data. Because the business project type is singular, the data it can collect is solely related to that business project, and therefore does not involve data from other feature providers. It should be noted that there can be one or more feature providers to enable the model training phase by combining features from different feature providers, thereby improving the model's generalization ability. Correspondingly, the label provider refers to the business entity that provides labels corresponding to the samples used in model training. The label provider can also provide additional data in other dimensions, which can be combined with the features from the feature providers during model training. The business entity corresponding to the label provider is also considered a business entity.

[0045] In practical implementation, when multiple business parties need to jointly train a model, the business parties providing the corresponding features and labels can be set according to actual needs. For example, business party A1 provides user transaction data, business party A2 provides user credit data, and business party A3 provides user online payment data. In a scenario where it is necessary to predict whether a user will activate online payment services, the ends corresponding to business parties A1 and A2 can be used as feature providers, and the end corresponding to business party A3 can be used as a label provider to train the prediction model for the current scenario. Alternatively, in a business scenario where it is necessary to predict whether a user will purchase a certain product, the ends corresponding to business parties A2 and A3 can be used as feature providers, and the end corresponding to business party A1 can be used as a label provider to train the prediction model for the current scenario. In practical applications, the settings for label providers and feature providers can be selected according to actual needs, and this embodiment does not impose any limitations.

[0046] Accordingly, the local encrypted model specifically refers to the local plaintext model of the feature provider, which is trained using encrypted data sent by the tag provider, and the model parameters are encrypted. This model can perform preliminary prediction processing based on the feature data of the feature provider's local data. That is, it can process local business features based on the local encrypted model to obtain the encrypted prediction result output by the model. Here, local business features specifically refer to the vector representation constructed based on the feature data provided by the business party to which the feature provider belongs. Correspondingly, the encrypted prediction result specifically refers to the encrypted prediction value obtained after processing the local business features using the local encrypted model. In other words, the prediction result is in an encrypted state to ensure security. At the same time, this result is the result obtained after processing by the feature provider's local encrypted model and is not the content corresponding to the original feature data. Therefore, it can further prevent other parties from deducing the original data content.

[0047] In practical applications, the predictive power of the local encrypted model is related to the type of feature data provided by the feature provider. Although the inputs to the encrypted models held by each feature provider are different, the output encrypted prediction results belong to the same type. For example, the predictive power of the models held by business parties A1 and A2 is to predict the level of a user's online payment service activation. The higher the level, the higher the probability of activation, and the lower the level, the lower the probability of activation. Business party A1 can input user transaction data into its encrypted model for processing to obtain the level of each user's online payment service activation. Business party A2 can input user credit data into its encrypted model for processing to obtain the level of each user's online payment service activation. This allows the prediction results to be aggregated at the label provider, and a business model that meets the current scenario can be trained based on the labels and feature data held by the label provider.

[0048] Furthermore, intermediate features specifically refer to the features obtained after decrypting the ciphertext prediction results. These features correspond to the prediction results of local business features. Although provided to the label provider, the label provider cannot deduce local business features from intermediate features because they have been processed by the model, and the model parameters are unknown to the label provider. This further ensures data security. Correspondingly, local associated features specifically refer to the feature data held by the business party corresponding to the label provider. Similarly, sample features specifically refer to the features used as input to the model for processing. They combine the intermediate features corresponding to the local business features of the feature provider, as well as the local associated features held by the label provider. Therefore, the sample features have richer dimensions, enabling the training of a business model with stronger generalization capabilities. Correspondingly, local sample labels specifically refer to the labels corresponding to the sample features, used to optimize the model during the training phase. Correspondingly, the business model specifically refers to the model that needs to be trained in the current business scenario, used to meet business requirements in the current scenario.

[0049] Accordingly, the training stopping condition specifically refers to the conditions for stopping the training of the business model, including but not limited to loss value comparison conditions or iteration count comparison conditions. The loss value comparison condition specifically refers to calculating the loss value by combining the model's output and labels, comparing the loss value with a preset loss value threshold, and determining whether the set condition is met based on the comparison result. If the condition is met, training stops; otherwise, training continues using sample pairs from the business sample set. The iteration count comparison condition specifically refers to comparing the number of iterations of model training with an iteration count threshold. During model training, training stops when the iteration count reaches the threshold. Correspondingly, the target business model specifically refers to a model that meets the training stopping condition and can be used in the current business scenario. This model can be a data model such as linear regression, logistic regression, or regression decision tree. In specific implementation, the model is selected according to the actual application scenario; this embodiment does not impose any limitations.

[0050] It's important to note that secure multi-party computation achieves data sharing by making data invisible to the participating parties. For example, if P0 and P1 are two parties involved in secure computation, we can let A be a finite abelian group. The sum-sharing form in G means that an element x in A exists in both P0 and P1 as x0 + x1. Here, x0 in A is stored in P0 and is invisible to P1, while x1 in A is stored in P1 and is invisible to P0. Furthermore, the addition in x0 + x1 refers to addition in the abelian group. This secure computation under the sum-sharing form requires P0 and P1 to compute the value of x in f: A->B through secure interactions, and then store this value in the sum-sharing form in both P0 and P1. Using this implementation, model training can be completed while ensuring data security, thus solving the data silo problem.

[0051] Based on this, taking secure matrix multiplication as an example, let x be an m x n matrix and y be an n x p matrix, stored in a shared manner on both P0 and P1. To calculate xy, we can use (x0 + x1)(y0 + y1) = (x0 + x1 - u0 - u1)(y0 + y1) + (u0 + u1)(y0 + y1 - v0 - v1) + (u0 + u1)(v0 + v1) to calculate the value of xy, supporting downstream business processing by both P0 and P1 based on the calculation result. Furthermore, according to the above calculation, the communication volume between P0 and P1 during matrix multiplication is 2(mn + np + mp) + 2(mn + np). If PRF is used, the communication volume becomes mp + 2(mn + np), which is also enormous. In the multi-party collaborative model training phase, offline training involves training the model jointly based on batch features and labels, while the online prediction phase involves training the model based on a single feature and the model's output prediction value. Both offline training and online prediction phases require a large number of arithmetic operations, logical operations, and comparison operations.

[0052] In view of this, see Figure 2 As shown, during the offline training phase, Pi and Q execute a modeling protocol. Pi creates a local encrypted model EQ(w, b), and then uses the local encrypted model EQ(w, b) to process local business features to obtain an encrypted prediction value EQ(Zi). This prediction value can be sent to Q with minimal communication. Q can then decrypt the encrypted prediction value EQ(Zi) to obtain intermediate features Zi. Combined with local associated features and labels, the upper-layer model M can be trained for use in the application phase.

[0053] Therefore, when training a model jointly by multiple parties, considering the various computations involved and the complexity of these computations leading to multiple communications and significant communication volume, which severely impacts training efficiency and data security, a solution is needed. To reduce communication volume while ensuring data security, the feature provider can process its local business features using a local encrypted model to obtain encrypted prediction results, which are then sent to the tag provider. The tag provider can decrypt the received encrypted prediction results to obtain intermediate features, then read local associated features and fuse them with the intermediate features to construct sample features. These features are then combined with local sample labels to build a business sample set containing numerous sample pairs. The tag provider can then use this business sample set to train a business model that meets the needs of the actual business scenario. This target business model can then be used by all parties involved in training the model during the application phase to complete prediction processing.

[0054] For ease of description, this embodiment uses a model trained with two feature providers and one label provider to predict whether a user has activated online payment services as an example to illustrate the above system. Applications in other scenarios can refer to the same or corresponding descriptions in this embodiment, and will not be elaborated further here.

[0055] Feature provider P1 holds users' fund flow data, feature provider P2 holds users' credit data, and tag provider Q holds users' transaction data. Transaction data includes both online and offline transactions, with online transaction data clearly indicating whether a user has activated online payment services. All data held by each party is considered user privacy data. Therefore, to train the business model without disclosing the data to other parties, each provider can perform initial predictions locally and then send encrypted data to tag provider Q for model training.

[0056] In other words, feature provider P1 can convert the fund flow data corresponding to users U1 to Un into feature matrices X1-[n, m1], and feature provider P2 can convert the credit data corresponding to users U1 to Un into feature matrices X2-[n, m2]. Then, feature provider P1 inputs the feature matrices X1 of the fund flow data corresponding to users U1 to Un into a local encrypted model for processing. This model is used to obtain the online payment service activation level for users U1 to Un, and the output of the encrypted model is always an encrypted prediction result, i.e., the encrypted prediction value EQ(Z1) corresponding to the activation level of online payment services for users U1 to Un. Simultaneously, feature provider P2 inputs the feature matrices X2 of the credit data corresponding to users U1 to Un into a local encrypted model for processing. This model is used to obtain the online payment service activation level for users U1 to Un, and the output of the encrypted model is always an encrypted prediction result, i.e., the encrypted prediction value EQ(Z2) corresponding to the activation level of online payment services for users U1 to Un. At this point, feature providers P1 and P2 can send the ciphertext prediction values ​​EQ(Z1) and EQ(Z2) to tag provider Q.

[0057] Furthermore, after receiving the encrypted prediction values ​​EQ(Z1) and EQ(Z2), the tag provider Q can decrypt them to obtain the intermediate feature Z1, after the feature provider P1's prediction processing, indicating the level at which users U1 to Un have activated online payment services; and the intermediate feature Z2, after the feature provider P2's prediction processing, indicating the level at which users U1 to Un have activated online payment services. Then, the tag provider Q can fuse the business feature Z0 corresponding to the offline transaction data of users U1 to Un with the intermediate features Z1 and Z2 to obtain the sample features Z corresponding to users U1 to Un respectively. Simultaneously, the tag provider Q will determine whether users U1 to Un have activated online payment services based on their online transaction data, and construct sample labels based on the determination result. After processing the sample features Z and sample labels according to user identity, n sample pairs can be obtained. These n sample pairs are then used to train the business model until a target business model that meets the training stopping condition is obtained. By utilizing the target business model, it is possible to predict whether a user will activate online payment services, so as to accurately target users for recommendations during the online payment service recommendation stage.

[0058] In summary, to complete model training while reducing communication volume, the feature provider can process local business features based on a local encrypted model, obtain encrypted prediction results, and send them to the label provider. The label provider can then determine intermediate features based on the encrypted prediction results, allowing the label provider to train the model using these intermediate features without disclosing data to the label provider. Furthermore, the label provider can construct sample features based on the intermediate features and their associated local features. Then, it can combine local sample labels with these sample features to form a business sample set. Using this business sample set to train the business model yields the target business model that meets the business scenario requirements. This process, where the feature provider uses a local encrypted model and the label provider combines encrypted prediction results from other endpoints to construct the sample set, completes business model training with minimal communication, effectively reducing communication volume while ensuring data security at all endpoints. This ensures that the trained target business model meets the needs of the business scenario.

[0059] Furthermore, to ensure data security while training a target business model that meets the current usage scenario, local encrypted models can be trained first for each feature provider. This allows for the subsequent training of the upper-layer model by combining the prediction results of the local encrypted models with the labels held by the feature providers. In this embodiment, the label provider is further configured to obtain encrypted business features by homomorphically encrypting the initial business features and sending them to the feature provider. The feature provider is also configured to receive the encrypted business features, construct an encrypted sample set based on the encrypted business features and local sample features, and train the initial local encrypted model using the encrypted sample set to obtain the local encrypted model. The model parameters of the local encrypted model are in an encrypted state.

[0060] Specifically, the initial business feature refers to the vector representation constructed from the feature data held by the tag provider. Correspondingly, the encrypted business feature refers to the vector representation encrypted by the tag provider using a key; it is a ciphertext feature relative to the feature provider and can only be decrypted by the tag provider. Correspondingly, the local sample feature refers to the vector representation constructed from the feature data held locally by the feature provider. Correspondingly, the ciphertext sample set refers to the set of sample pairs constructed by combining the encrypted initial business feature and the local sample feature; it is used to train the initial local ciphertext model of the feature provider to obtain the local ciphertext model held by the feature provider. The parameters of the local ciphertext model are encrypted and can only be decrypted by the tag provider.

[0061] Therefore, when the feature provider and label provider jointly train the model, considering that the data held by the feature provider is relatively limited and the business model to be trained in the current business scenario cannot be trained solely using the data provided by the feature provider, the label provider can first encrypt the initial business features held locally using homomorphic encryption to obtain encrypted business feature data before sending it to the feature provider.

[0062] Furthermore, after receiving the encrypted business features, the feature provider can combine local sample features and encrypted business features to construct sample pairs, and then integrate all sample pairs to form a ciphertext sample set. The initial ciphertext model can then be trained using this ciphertext sample set to obtain a local ciphertext model that meets the training stopping condition. During this process, because the sample pairs used in model training contain encrypted business features, the model parameters after training will also be in an encrypted state, making the model only usable relative to the feature provider, thus facilitating its use during the business model training phase.

[0063] The above process is described using a linear regression model as an example. Assume that P has a feature matrix X[n,m] and Q has labels Y[n,1]. First, the desired linear regression model is... w∈R m For b∈R, during training, firstly, Q-side encrypts the label Y using homomorphic encryption, obtaining the encrypted result EQ(Y), and sends it to P-side. After receiving the encrypted result EQ(Y), P-side calculates the model parameters using plaintext-ciphertext multiplication and ciphertext addition, obtaining... Since the model parameters are calculated jointly by combining the encrypted EQ(Y) and the local features of P, therefore Since P is a parameter of the encryption model, the local ciphertext model held by P can be obtained at this point for convenient use later.

[0064] It should also be noted that if Q needs to use P's model, P can send the encrypted model to Q, and Q can decrypt the model using its local key to obtain the plaintext model for use.

[0065] It should be noted that since there are multiple feature providers, each feature provider implements the above description when training its local encrypted model, so that each feature provider can have a local encrypted model, which facilitates subsequent joint training of the business model by multiple parties.

[0066] In summary, by using homomorphic encryption to send encrypted business features to the feature provider, the feature provider can train a local ciphertext model by combining the encrypted business features and local sample features. This facilitates the subsequent training of the business model using the local ciphertext model. Furthermore, the communication between the feature provider and the tag provider occurs only once during this process, effectively reducing communication volume and further improving model training efficiency.

[0067] Furthermore, since the data held by each party is private data relative to other parties, it can be ensured that both parties can complete model creation and training with less communication. It can also be completed according to the modeling protocol between the two parties. In this embodiment, the feature providing end 110 is also configured to create the local encrypted model by executing the modeling protocol associated with the tag providing end; the tag providing end 120 is also configured to create the business model by executing the modeling protocol associated with the feature providing end.

[0068] Specifically, the modeling protocol associated with the tag provider is the same as that associated with the feature provider, but the models created based on the modeling protocol are different. The local encrypted model created by the feature provider can complete the preliminary prediction based on local data, while the business model created by the tag provider can integrate the encrypted prediction results of various feature providers to complete the final prediction to meet the needs of the scenario.

[0069] Therefore, when a business model that meets the current scenario needs to be trained, each feature provider and label provider can execute a modeling protocol to build its own model. In other words, the feature provider will build a local encrypted model, while the label provider will build a business model for use in subsequent model training phases.

[0070] It should be noted that after the business model is established on the label provider side, it needs to be combined with other feature providers to complete the training. The feature providers can train their local encrypted models based on their locally held data. Although the trained models may have overfitting issues, the output of these models is not the final result for the application stage. It still needs to be processed in conjunction with the target business model trained by the label providers to ensure prediction accuracy.

[0071] In summary, by adopting an execution modeling protocol, the feature provider and label provider can build models first to support subsequent training of business models, which can effectively reduce the amount of communication between them and improve model training efficiency.

[0072] Furthermore, during the modeling process, to prevent the tag provider from deriving the original data based on the ciphertext prediction results, which could lead to the risk of privacy data leakage, the feature provider can encrypt the model parameters after modeling and use the ciphertext model to complete the prediction. The prediction result is also in ciphertext, further ensuring data security. In this embodiment, the feature provider 110 is also configured to create a local plaintext model by executing the modeling protocol associated with the tag provider, and use a local key to perform homomorphic encryption on the model parameters of the local plaintext model to obtain the local ciphertext model.

[0073] Specifically, the local plaintext model refers to a model created by the feature provider according to the modeling protocol associated with the tag provider, where the model parameters are not encrypted. Correspondingly, the local key refers to a key held by the feature provider, which is invisible to other feature providers and tag providers, ensuring that the encrypted model parameters cannot be cracked or obtained.

[0074] Therefore, in order to ensure the security of the model during use and to avoid affecting the prediction results, a local plaintext model can be created by executing the modeling protocol first. Then, the model parameters of the local plaintext model can be homomorphically encrypted using a local key to obtain a local ciphertext model. During the application phase, the local ciphertext model can be used to process local business features and output ciphertext prediction results, so that other feature providers and tag providers cannot deduce the parameters of the model and the feature data input to the model.

[0075] It should be noted that after the local plaintext model is created, it can be trained using local business features until a local plaintext model that meets the training stopping condition is obtained. Then, the model parameters are homomorphically encrypted to obtain the local ciphertext model held by the feature provider itself.

[0076] Continuing with the previous example, during the training phase, feature provider P1, feature provider P2, and label provider Q can complete model creation by executing a three-party modeling protocol. To ensure data security, feature provider P1 can first establish a plaintext model [EQ1(w1), EQ1(b1)], and feature provider P2 can first establish a plaintext model [EQ2(w2), EQ2(b2)]. Simultaneously, label provider Q will establish a business model. Furthermore, feature provider P1 uses its local key to perform homomorphic encryption on the plaintext model [EQ1(w1), EQ1(b1)] to obtain the ciphertext model [EQ(w1), EQ(b1)]. Feature provider P2 uses its local key to perform homomorphic encryption on the plaintext model [EQ2(w2), EQ2(b2)] to obtain the ciphertext model [EQ(w2), EQ(b2)].

[0077] During the training phase of the business model, feature provider P1 can calculate the encrypted prediction value EQ(Z1) = X1 EQ(w1) + EQ(b1) for its own fund flow data X1, and feature provider P2 can calculate the encrypted prediction value EQ(Z2) = X2 EQ(w2) + EQ(b2) for its own credit data X2. Then, the encrypted prediction values ​​are sent to tag provider Q for processing.

[0078] In summary, by encrypting the model parameters using homomorphic encryption, the model parameters can be made unknowable, and the local business characteristics of the input model cannot be derived from the model's prediction results, thereby further improving data security.

[0079] During the business model training phase at the label provider, since the label provider integrates the encrypted prediction results from all feature providers, in order to support the label provider in constructing sample features that meet the model training requirements, the label provider needs to perform decryption processing before adjustment and fusion. In this embodiment, the label provider 120 is further configured to receive the encrypted prediction results, decrypt the encrypted prediction results to obtain the intermediate features, read the local associated features, and integrate the intermediate features and the local associated features to obtain the sample features.

[0080] Based on this, after the tag provider receives the encrypted prediction results from each feature provider, it can first read the public key of each feature provider locally, and then use the public key to decrypt the encrypted prediction results to obtain the intermediate features corresponding to each feature provider. Then, it can read the local associated features related to the business scenario locally, and fuse the local associated features with the intermediate features to obtain the sample features, which can then be used during the model training phase.

[0081] In summary, by sending encrypted prediction results to the tag provider, data security can be achieved. At the same time, the tag provider can use the public key to decrypt the data, enabling them to construct sample features based on the decryption results, thereby improving the generalization ability of the model.

[0082] Based on this, when training the model, the label provider needs to clarify that intermediate features and locally associated features correspond to each other by combining the relationships between features, so that the constructed sample features are usable and can be used to train a more accurate business model. In this embodiment, the label provider 120 is further configured to determine the feature identification information corresponding to the intermediate features and the associated feature identification information corresponding to the locally associated features, establish feature association relationships based on the feature identification information and the associated feature identification information, and integrate the intermediate features and the locally associated features according to the feature association relationships to obtain the sample features.

[0083] Specifically, feature identification information refers to the identification information that clearly identifies the object to which the intermediate feature belongs. For example, in online payment scenarios, the intermediate feature corresponds to the user's level of online payment service activation, and the feature identification information is the user's identity identifier; in medical scenarios, the intermediate feature corresponds to the probability of a user having disease A, and the feature identification information is the name of disease A; in shopping scenarios, the intermediate feature corresponds to the probability of a user purchasing product B, and the feature identification information is the name of product B's brand. The feature identification information differs in different scenarios. Correspondingly, associated feature identification information refers to the identification information corresponding to the associated features held by the tag provider, and it has the same concept as feature identification information. Correspondingly, feature association specifically refers to the similarity between feature identification information and associated feature identification, and based on this association, intermediate features and locally associated features can be fused.

[0084] Based on this, after obtaining intermediate features from the label provider, to ensure that intermediate features and locally associated features can be fused, and that the fused sample features belong to the same object, we can first determine the feature identifier information corresponding to the intermediate features, as well as the associated feature identifier information of the locally associated features. The feature identifier information can be carried by the intermediate features and written into the encrypted prediction result by the feature provider. After obtaining the feature identifier information and associated feature identifier information, we can select the same identifier information to establish feature association relationships. Then, based on these feature association relationships, we can integrate the intermediate features and locally associated features of the same object, thus fusing them together to obtain the sample features of each object. These features can then be used for subsequent model training.

[0085] In summary, by determining feature identification information and establishing feature associations, we can identify intermediate features and local associated features belonging to the same object. Then, we can fuse features according to feature associations, and the constructed sample features all belong to each object. This is used to accurately construct sample features, thereby enabling the training of a business model with higher accuracy and stronger generalization ability.

[0086] Furthermore, considering that the business sample set is a collection constructed based on sample features and local sample labels, when training the business model, samples can be selected sequentially from the set to train the model, and after each prediction process, parameters are tuned in conjunction with the labels. In this embodiment, the label providing end 120 is further configured to input the sample features from the business sample set into the business model for processing to obtain sample prediction information, and to tune the parameters of the business model based on the sample prediction information and the local sample labels from the business sample set until the tuning result meets the training stopping condition to obtain the target business model.

[0087] Based on this, after constructing the sample features, local sample labels associated with the sample features can be determined according to the local feature identification information. Then, multiple sample pairs are formed by combining the sample features and local sample labels. Integrating all sample pairs yields the business sample set. During the training phase, sample features can be extracted from the business sample set and input into the business model for processing to obtain sample prediction information. Then, local sample labels corresponding to the sample features can be extracted from the business sample set. The loss value is calculated using the local sample labels and sample prediction information. The parameters of the business model are tuned based on the loss value. This process is repeated until the target business model that meets the training stopping condition is obtained.

[0088] In practical applications, during model training, to improve the prediction accuracy of the trained model, a validation set can be constructed by combining sample features and local sample labels. That is, after each iteration of training, the sample features from the validation set can be input into the model to obtain its output. The output is then compared with the labels in the validation set. If they do not match, it indicates that the model's prediction accuracy can be improved, and training can continue using the business sample set. Training stops when the model's output matches the labels using the validation set, and this stage of the model becomes the target business model. It should be noted that to avoid randomness, the validation stage can combine multiple sample features for verification.

[0089] For example, if user U1's online payment service activation level is determined to be L1 and user U2's is L2, then the tag provider Q determines user U1's identifier as ID_U1 and user U2's identifier as ID_U2. In the local offline transaction data, offline transaction data S1 corresponds to identifier ID_U1, and offline transaction data S2 corresponds to identifier ID_U2. Therefore, based on the similarity of the identifier information, user U1's online payment service activation level can be merged with offline transaction data S1 to obtain the sample features corresponding to user U1. Similarly, user U2's online payment service activation level can be merged with offline transaction data S2 to obtain the sample features corresponding to user U2.

[0090] This process continues until all user sample features are constructed. Then, the results of each user activating online payment services can be used as sample labels. These labels and features are then matched according to user identity, resulting in a sample set containing n sample pairs. During training, the sample features of any user in the set can be selected and input into the business model for processing to obtain the probability of that user activating online payment services. The loss value is then calculated using the probability and its corresponding sample label. Model parameters are tuned according to the loss value until the target business model that meets the training stopping condition is obtained. This allows the target business model to predict whether a user will activate online payment services, enabling accurate user recommendations during the online payment service recommendation phase.

[0091] In summary, training the business model using a business sample set can effectively improve the model training accuracy. Furthermore, the business model serves as the top-level model, and the training is completed using a stacking approach, which further ensures data security. At the same time, no additional communication is required, effectively saving transmission resources.

[0092] Furthermore, after the model training is completed, considering that all participants may use the model, in order to facilitate the use of the target business model by all parties, the tag provider 120 is also configured to broadcast the target business model to the feature provider; or, the model parameters of the target business model are homomorphically encrypted using the local tag key to obtain the target business encrypted model, the target business encrypted model is stored in the target location, and the target location is broadcast to the feature provider.

[0093] Therefore, to avoid privacy leaks when broadcasting the target business model to various feature providers, the tag provider can first perform homomorphic encryption on the encrypted business model before broadcasting it. Alternatively, it can use a local tag key to perform homomorphic encryption on the model parameters of the target business model to obtain the encrypted target business model. This encrypted model can then be stored at the target location and broadcast to the feature providers. This allows the feature providers to retrieve the encrypted target business model from the target location as needed.

[0094] In summary, by informing each feature provider that the model training is complete, it becomes easier for the feature providers to use the target business model based on the broadcast results during the application phase, so as to complete the usage processing in the current business scenario.

[0095] Figure 3This is a schematic diagram of a service processing system based on multi-party secure computation according to one embodiment of this specification. The service processing system 300 based on multi-party secure computation includes a feature providing end 310 and a tag providing end 320. The system includes:

[0096] The tag provider 320 is configured to acquire business prediction information, create a business prediction request based on the business prediction information, and send it to the feature provider.

[0097] The feature provider 310 is configured to read the target business features in response to the business prediction request, process the target business features using a local ciphertext model, obtain the target ciphertext prediction result, and send it to the tag provider.

[0098] The tag provider 320 is also configured to receive the target ciphertext prediction result, determine the target intermediate features based on the target ciphertext prediction result, construct target features based on the target intermediate features and target local association features, and input the target features into the target business model in the above system for processing to obtain business prediction results.

[0099] The service processing system based on multi-party secure computation provided in this embodiment corresponds to the service model training method based on multi-party secure computation described above. Similar or identical descriptions can be found in the above embodiments, and will not be elaborated upon further here. It should be noted that the feature provider 310 and tag provider 320 included in the service processing system 300 based on multi-party secure computation provided in this embodiment are designated during the prediction phase and are the same as the feature provider and tag provider in the above system.

[0100] Based on this, since the feature provider maintains a local encrypted model, while the tag provider maintains a top-level model—the target business model—and the input to the target business model is the intermediate features corresponding to the encrypted prediction results of each feature provider, during the online prediction phase, after obtaining business prediction information, the tag provider can first create a business prediction request and send it to the feature provider. The feature provider then responds to the business prediction request by reading the locally stored target business features, processing them using the local encrypted model, and obtaining the target encrypted prediction result, which is then sent to the tag provider. This allows the tag provider to combine relevant features from the feature provider when performing business prediction processing.

[0101] Furthermore, after receiving the target encrypted prediction result, the tag provider can decrypt it to obtain the target intermediate features. Then, it can combine the target local association features stored locally to construct the target features, and then input them into the target business model trained by the above system for processing to obtain the business prediction result of the corresponding business prediction information.

[0102] Furthermore, to ensure that all features used in the prediction phase are associated with the same object, the prediction processing operation can be completed by combining the identification information. In this embodiment, the tag provider 320 is further configured to parse the business prediction information to obtain business identification information, create the business prediction request based on the business identification information, and send it to the feature provider; the feature provider 310 is further configured to determine the business identification information in response to the business prediction request, and read the business feature associated with the business identification information as the target business feature.

[0103] Based on this, after receiving business prediction information, the tag provider can parse it to obtain business identification information, and use this as a basis to create a business prediction request and send it to the feature provider. The feature provider can then accurately read the associated business features locally based on the business identification information as target business features for subsequent prediction processing.

[0104] In other words, see Figure 2 As shown, after obtaining the business identification information ID, the tag provider Q sends it to the feature provider Pi. The feature provider Pi first reads the target business feature corresponding to the business identification information ID, and then processes it using the local encrypted model EQ(w, b) to obtain the target encrypted prediction result EQ(Zi). Then, it feeds it back to the tag provider Q. The tag provider Q obtains Zi by decrypting all the target encrypted prediction results EQ(Zi) from the feature providers. Then, it combines the local associated features to form the target feature, and finally inputs it into the upper-layer model M to obtain the prediction result.

[0105] For example, if tag provider Q receives user Um's identifier information as ID_Um, it can send ID_Um to feature providers A1 and A2. Feature provider A1 will read user Um's fund transfer data based on ID_Um, and feature provider A2 will read user Um's credit data based on ID_Um. Both feature providers will then process this data using their local encrypted model. Feature provider A1 will then send back an encrypted prediction value Y1 (user Um's level of online payment service activation) to tag provider Q based on the processing result. Feature provider A2 will also send back an encrypted prediction value Y2 (user Um's level of online payment service activation) to tag provider Q based on the processing result. After receiving the encrypted prediction values ​​Y1 and Y2, tag provider Q can combine them with user Um's offline payment data held locally to form a target feature. Finally, this feature is input into the target business model for processing, thus obtaining the probability that user Um has activated online payment services. Based on this probability, it can then decide whether to send recommendation information to the user.

[0106] The business processing system based on multi-party secure computation provided in this specification, after obtaining business prediction information at the tag provider, can create a business prediction request based on the business prediction information and send it to the feature provider that stores associated business prediction request data. The feature provider can respond to the business prediction request by reading the target business features, processing them using a local encrypted model, obtaining the target encrypted prediction result, and then sending it to the tag provider. This method of obtaining encrypted data through encrypted model processing and sending it to the tag provider allows the tag provider to use the features it needs without disclosing the data, effectively ensuring data security. Subsequently, the tag provider can generate target intermediate features based on the target encrypted prediction result, and construct target features by combining them with the target local associated features. By processing the target business model, the corresponding business prediction result can be obtained. This system enables the joint processing of prediction data from multiple parties, which not only ensures security but also reduces the number of communications between multiple terminals, thereby reducing the consumption of transmission resources.

[0107] The following is in conjunction with the appendix Figure 4 Taking the application of the business processing system based on multi-party secure computation provided in this specification in a shopping application scenario as an example, the business processing system based on multi-party secure computation will be further explained. Among them, Figure 4 This specification illustrates a schematic diagram of a business processing system based on multi-party secure computation according to an embodiment of this specification; specifically, it includes the following steps.

[0108] In step S402, the feature provider creates a local ciphertext model by executing the modeling protocol associated with the tag provider.

[0109] Step S404: The feature provider processes the local business features based on the local ciphertext model to obtain the ciphertext prediction result.

[0110] Step S406: The feature provider sends the ciphertext prediction result to the label provider.

[0111] In step S408, the label provider decrypts the ciphertext prediction result to obtain intermediate features.

[0112] In step S410, the label provider determines the feature identifier information corresponding to the intermediate feature and the associated feature identifier information corresponding to the local associated feature.

[0113] In step S412, the tag provider establishes a feature association relationship based on the feature identification information and the associated feature identification information.

[0114] In step S414, the label provider integrates intermediate features and local associated features according to feature association relationships to obtain sample features.

[0115] Step S416: The tag provider assembles a business sample set based on the sample features and local sample tags.

[0116] In step S418, the tag provider uses the business sample set to train the business model until a target business model that meets the training stopping condition is obtained.

[0117] Step S420: The tag provider receives service identification information.

[0118] Step S422: The tag provider sends the service identification information to the feature provider.

[0119] Step S424: The feature provider reads the business feature associated with the business identifier information as the target business feature based on the business identifier information.

[0120] Step S426: The feature provider processes the target business features using the local ciphertext model to obtain the target ciphertext prediction result.

[0121] Step S428: The feature provider sends the target ciphertext prediction result to the tag provider.

[0122] In step S430, the tag provider determines the intermediate features of the target based on the target ciphertext prediction results, and constructs target features based on the intermediate features and the target local association features.

[0123] In step S432, the label provider inputs the target features into the target business model for processing to obtain the business prediction result.

[0124] In summary, during the model training phase, the feature provider uses a local encrypted model, while the label provider combines encrypted prediction results from other sources to construct a sample set. This approach completes the training of the business model with minimal communication, effectively reducing communication volume while ensuring data security at each end. This allows the trained target business model to meet the needs of the business scenario. Once the model training is complete, if prediction of information within the current business scenario is required, it can be accomplished by combining data from multiple sources. However, data is not directly cross-referenced between these sources; instead, it is processed locally and then delivered to the label provider. This not only ensures security but also reduces the number of communications between multiple sources, effectively minimizing transmission resource consumption.

[0125] Corresponding to the above method embodiments, this specification also provides embodiments of a business model training method based on multi-party secure computation. Figure 5 This is a flowchart illustrating a business model training method based on multi-party secure computation, provided in one embodiment of this specification. Figure 5 As shown, the method includes:

[0126] Step S502: Process the local business features based on the local ciphertext model to obtain the ciphertext prediction result;

[0127] Step S504: Determine intermediate features based on the ciphertext prediction results, construct sample features based on the intermediate features and local association features, and form a business sample set based on the sample features and local sample labels. Use the business sample set to train the business model until a target business model that meets the training stopping condition is obtained.

[0128] In one optional embodiment, the tag provider obtains encrypted service features by performing homomorphic encryption on the initial service features, and sends them to the feature provider.

[0129] The feature provider receives the encrypted service features, constructs a ciphertext sample set based on the encrypted service features and local sample features, and trains the initial local ciphertext model using the ciphertext sample set to obtain the local ciphertext model; wherein, the model parameters of the local ciphertext model are the encryption states.

[0130] In one optional embodiment, the tag provider receives the ciphertext prediction result, decrypts the ciphertext prediction result to obtain the intermediate feature, reads the local association feature, and integrates the intermediate feature and the local association feature to obtain the sample feature.

[0131] In an optional embodiment, the label provider determines the feature identifier information corresponding to the intermediate feature and the associated feature identifier information corresponding to the local associated feature, establishes a feature association relationship based on the feature identifier information and the associated feature identifier information, and integrates the intermediate feature and the local associated feature according to the feature association relationship to obtain the sample feature.

[0132] In one optional embodiment, the label provider inputs the sample features from the business sample set into the business model for processing to obtain sample prediction information. Based on the sample prediction information and the local sample labels in the business sample set, the business model is tuned until the tuning result meets the training stopping condition, thereby obtaining the target business model.

[0133] In one optional embodiment, the tag provider broadcasts the target service model to the feature provider; or, the model parameters of the target service model are homomorphically encrypted using a local tag key to obtain a target service ciphertext model, the target service ciphertext model is stored in the target location, and the target location is broadcast to the feature provider.

[0134] In one optional embodiment, the feature provider creates the local ciphertext model by executing a modeling protocol associated with the tag provider;

[0135] The tag provider creates the business model by executing a modeling protocol associated with the feature provider.

[0136] In one optional embodiment, the feature provider creates a local plaintext model by executing a modeling protocol associated with the tag provider, and uses a local key to perform homomorphic encryption on the model parameters of the local plaintext model to obtain the local ciphertext model.

[0137] The business model training method based on multi-party secure computation provided in this specification aims to complete model training while reducing communication volume. The feature provider can process local business features based on a local encrypted model, obtain encrypted prediction results, and send them to the tag provider. The tag provider can then determine intermediate features based on the encrypted prediction results, enabling the tag provider to train the model using these intermediate features without disclosing data to the tag provider. Furthermore, the tag provider can construct sample features based on the intermediate features and their associated local features. Then, it combines local sample labels with these sample features to form a business sample set. Using this business sample set to train the business model yields a target business model that meets the business scenario requirements. This process, where the feature provider uses a local encrypted model and the tag provider combines encrypted prediction results from other sources to construct the sample set, completes business model training with minimal communication, effectively reducing communication volume while ensuring data security at all ends. This ensures that the trained target business model meets the needs of the business scenario.

[0138] The above is an illustrative scheme of a business model training method based on multi-party secure computation according to this embodiment. It should be noted that the technical solution of this business model training method based on multi-party secure computation belongs to the same concept as the technical solution of the business model training system based on multi-party secure computation described above. For details not described in detail in the technical solution of the business model training method based on multi-party secure computation, please refer to the description of the technical solution of the business model training system based on multi-party secure computation described above.

[0139] Corresponding to the above method embodiments, this specification also provides embodiments of a business processing method based on multi-party secure computation. Figure 6 This is a flowchart illustrating a business processing method based on multi-party secure computation, as provided in one embodiment of this specification. Figure 6 As shown, the method includes:

[0140] Step S602: Obtain business forecast information and create a business forecast request based on the business forecast information;

[0141] Step S604: In response to the business prediction request, read the target business features, process the target business features using the local encrypted model, and obtain the target encrypted prediction result;

[0142] Step S606: Determine the target intermediate features based on the target ciphertext prediction result, construct the target features based on the target intermediate features and the target local association features, and input the target features into the target business model in the above method for processing to obtain the business prediction result.

[0143] In one optional embodiment, the tag provider parses the business prediction information to obtain business identification information, and creates the business prediction request based on the business identification information;

[0144] The feature provider responds to the service prediction request to determine the service identification information and reads the service features associated with the service identification information as the target service feature.

[0145] The business processing method based on multi-party secure computation provided in this specification allows the tag provider to create a business prediction request after obtaining business prediction information. This request is then sent to a feature provider that stores associated business prediction request data. The feature provider can respond to the business prediction request by reading the target business features, processing them using a local encrypted model, and then sending the encrypted prediction result to the tag provider. This method of obtaining encrypted data through encrypted model processing allows the tag provider to use the features it needs without disclosing the data, effectively ensuring data security. Subsequently, the tag provider can generate target intermediate features based on the encrypted prediction result, construct target features by combining them with the target local associated features, and process them using the target business model to obtain the corresponding business prediction result. This method enables the joint processing of prediction data from multiple parties, ensuring security and reducing the number of communications between multiple terminals, thereby reducing the consumption of transmission resources.

[0146] The above is an illustrative scheme of a business processing method based on multi-party secure computation according to this embodiment. It should be noted that the technical solution of this business processing method based on multi-party secure computation belongs to the same concept as the technical solution of the business processing system based on multi-party secure computation described above. For details not described in detail in the technical solution of the business processing method based on multi-party secure computation, please refer to the description of the technical solution of the business processing system based on multi-party secure computation described above.

[0147] Figure 7A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0148] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0149] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0150] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 700 can also be a mobile or stationary server.

[0151] The processor 720 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-mentioned business model training method or business processing method based on multi-party secure computation.

[0152] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the above-described business model training method or business processing method based on multi-party secure computation. For details not described in detail in the technical solution of the computing device, please refer to the description of the above-described business model training method or business processing method based on multi-party secure computation.

[0153] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described business model training method or business processing method based on multi-party secure computation.

[0154] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solutions of the aforementioned business model training method based on multi-party secure computation or business processing method based on multi-party secure computation. Details not described in detail in the technical solution of the storage medium can be found in the descriptions of the aforementioned business model training method based on multi-party secure computation or business processing method based on multi-party secure computation.

[0155] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described business model training method based on multi-party secure computation or business processing method based on multi-party secure computation.

[0156] The above is an illustrative example of a computer program in this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the aforementioned technical solutions of the business model training method or business processing method based on multi-party secure computation. Details not described in detail in the computer program's technical solution can be found in the descriptions of the aforementioned technical solutions of the business model training method or business processing method based on multi-party secure computation.

[0157] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0158] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0159] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0160] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0161] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A business model training system based on multi-party secure computation, comprising: The feature provider is configured to process local business features based on a local ciphertext model, obtain ciphertext prediction results, and send them to the tag provider. The prediction capability of the local ciphertext model is related to the feature data type of the feature provider, and the ciphertext prediction results are in an encrypted state. The tag provider is configured to receive the ciphertext prediction result, determine intermediate features based on the ciphertext prediction result, construct sample features based on the intermediate features and local association features, and form a business sample set based on the sample features and local sample tags. The business model is trained using the business sample set until a target business model that meets the training stopping condition is obtained. The intermediate features are features obtained after decrypting the ciphertext prediction result, and the intermediate features and the local association features belong to the same object.

2. The system according to claim 1, wherein the tag providing end is further configured to obtain encrypted service features by performing homomorphic encryption processing on the initial service features, and send them to the feature providing end; The feature provider is further configured to receive the encrypted service features, construct a ciphertext sample set based on the encrypted service features and local sample features, and train an initial local ciphertext model using the ciphertext sample set to obtain the local ciphertext model; wherein... The model parameters of the local ciphertext model are the encryption state.

3. The system according to claim 1, wherein the tag providing end is further configured to receive the ciphertext prediction result, decrypt the ciphertext prediction result to obtain the intermediate feature, read the local association feature, and integrate the intermediate feature and the local association feature to obtain the sample feature.

4. The system according to claim 3, wherein the tag providing end is further configured to determine the feature identification information corresponding to the intermediate feature and the associated feature identification information corresponding to the local associated feature, establish a feature association relationship according to the feature identification information and the associated feature identification information, and integrate the intermediate feature and the local associated feature according to the feature association relationship to obtain the sample feature.

5. The system according to claim 1, wherein the tag providing end is further configured to input the sample features of the business sample set into the business model for processing to obtain sample prediction information, and to tune the parameters of the business model according to the sample prediction information and the local sample tags of the business sample set until the tuning result meets the training stopping condition to obtain the target business model.

6. The system according to any one of claims 1-5, wherein the tag provider is further configured to broadcast the target service model to the feature provider; or, The model parameters of the target service model are homomorphically encrypted using a local tag key to obtain a target service ciphertext model. The target service ciphertext model is then stored at the target location, and the target location is broadcast to the feature provider.

7. The system of claim 1, wherein the feature provider is further configured to create the local ciphertext model by executing a modeling protocol associated with the tag provider; The tag provider is also configured to create the business model by executing a modeling protocol associated with the feature provider.

8. The system according to claim 7, wherein the feature provider is further configured to create a local plaintext model by executing a modeling protocol associated with the tag provider, and to perform homomorphic encryption on the model parameters of the local plaintext model using a local key to obtain the local ciphertext model.

9. A business processing system based on multi-party secure computation, comprising: The tag provider is configured to acquire business prediction information, create a business prediction request based on the business prediction information, and send it to the feature provider. The feature provider is configured to read the target business features in response to the business prediction request, process the target business features using a local encrypted model, obtain the target encrypted prediction result, and send it to the tag provider. The tag provider is further configured to receive the target ciphertext prediction result, determine the target intermediate features based on the target ciphertext prediction result, construct target features based on the target intermediate features and target local association features, and input the target features into the target business model in any one of claims 1-8 for processing to obtain business prediction results.

10. The system according to claim 9, wherein the tag provider is further configured to parse the service prediction information to obtain service identification information, create the service prediction request based on the service identification information, and send it to the feature provider; The feature provider is further configured to determine the service identification information in response to the service prediction request, and read the service features associated with the service identification information as the target service feature.

11. A business model training method based on multi-party secure computation, comprising: The local business features are processed based on the local ciphertext model to obtain ciphertext prediction results. The prediction capability of the local ciphertext model is related to the feature data type of the feature provider. The ciphertext prediction results are in an encrypted state. Intermediate features are determined based on the ciphertext prediction results. Sample features are constructed based on the intermediate features and local association features. A business sample set is formed based on the sample features and local sample labels. The business model is trained using the business sample set until a target business model that meets the training stopping condition is obtained. The intermediate features are features obtained after decrypting the ciphertext prediction results. The intermediate features and the local association features belong to the same object.

12. A business processing method based on multi-party secure computation, comprising: Obtain business forecast information and create a business forecast request based on the business forecast information; In response to the business prediction request, the target business features are read, and the target business features are processed using a local encrypted model to obtain the target encrypted prediction result. Based on the target ciphertext prediction result, target intermediate features are determined, and target features are constructed based on the target intermediate features and target local association features. The target features are then input into the target business model in the method of claim 10 for processing to obtain business prediction results.

13. A computing device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method of claim 11 or 12.

14. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method of claim 11 or 12.

15. A computer program product comprising computer instructions that, when executed by a processor, implement the steps of the method of claim 11 or 12.

Images