Off-site distributed security data transaction method for privacy protection

Abstract

The invention discloses an off-site distributed security data transaction method for privacy protection, and belongs to the field of privacy computing. According to the method, around direct transaction between a data seller and a data buyer, the problems of data tampering and data leakage possibly caused in a traditional centralized intermediary mode are solved, and operation autonomy and data transaction flexibility are returned to a data owner. Specifically, a security calculation method is innovatively introduced, and the balance of data security and availability between a data seller and a data buyer is realized. A security data analysis service provided for a buyer by a seller is presented through two specific security transaction examples (1) based on homomorphic encryption linear regression training and 2) based on secret sharing linear regression training, so that security transaction of data of the seller / buyer is ensured. According to the method, the problems of difficult data copying prevention and difficult secondary transaction prevention are effectively solved, the confidentiality and integrity of data of all parties in the transaction process are ensured, the risk of different credibility of participants is effectively handled, the interference of malicious participants is prevented, and the security and reliability of data transaction are further improved.

Description

Technical Field

[0001] The present invention relates to the field of privacy computing and provides an off-site decentralized secure data trading method with privacy protection. Background Art

[0002] With the rapid development of technologies such as big data, cloud computing and artificial intelligence, data has become an important resource for promoting social progress and industrial transformation. In this context, data trading, as the core means of releasing data value elements, has gradually developed into two main models: on-site data trading (centralized trading model) and off-site data trading (decentralized trading model).

[0003] In the on-site data trading model, data exchange relies on centralized servers for data storage and processing, which has the advantages of high efficiency and centralized management. However, this centralized architecture makes the attack point concentrated on a single platform, which is prone to single point failure. In addition, dishonest sellers or buyers may manipulate the transaction process and undermine the fairness of the platform, while untrusted servers may privately copy the original data during the data exchange process, causing the risk of data leakage or secondary transactions.

[0004] In contrast, the OTC data trading model breaks the limitations of a single platform through a decentralized approach, and data sellers and buyers can complete transactions through a decentralized network or multiple parties. This model has obvious advantages in improving transaction flexibility and achieving trustlessness, but it also brings new challenges. Since the security and credibility of the participants are difficult to guarantee, data leakage or tampering has become a major hidden danger. In addition, the participants are distributed in different network environments, and data is also vulnerable to various attacks when transmitted in untrusted networks.

[0005] How to ensure the confidentiality and integrity of data of all parties during the data transaction process and prevent untrusted parties from tampering with data, illegally copying or conducting secondary transactions has become a key issue that needs to be urgently addressed in the current data trading market. Summary of the invention

[0006] The present invention aims to solve the problems faced in existing data trading scenarios, such as difficulty in preventing data duplication and secondary transactions, as well as the theft of sensitive data by untrusted participants. At the same time, it overcomes the drawbacks of traditional centralized intermediary institutions and proposes a privacy-preserving over-the-counter decentralized secure data trading method that takes participants as the main body and realizes data exchange.

[0007] Specifically, this method bypasses traditional intermediary brokers (i.e., third parties responsible for centralized data processing) and relies solely on data sellers and data buyers to complete transactions directly. As two independent computing entities, data sellers and data buyers will complete data processing under privacy protection by collaboratively executing two-party secure computing operations (e.g., based on homomorphic encryption, secret sharing, obfuscated circuits, etc.) such as model training, reasoning, and statistical analysis. Under this architecture, either the data seller or the data buyer holds private inputs (e.g., the data seller provides a data set, and the data buyer provides a training task or an initial model to be trained). Under the constraints of a given function (i.e., the prescribed operation steps), both parties are allowed to jointly calculate their respective private inputs and obtain the calculation results. In this process, it is ensured that neither party can obtain the other party's private input.

[0008] This step can be formally expressed as: in is the encrypted data set provided by the seller, f is the established function provided by the buyer (the function is irreversible), and This is the return result obtained by the data buyer and does not include the original data X, thereby ensuring that only the right to use the data is traded, not the ownership of the data.

[0009] Furthermore, the secure transaction requires two-way security for data sellers and data buyers: on the one hand, data buyers can only obtain the results returned by the calculation, but cannot obtain the original data belonging to the data seller; on the other hand, data sellers cannot understand the intermediate calculation results and the private input provided by the data buyer (for example, the data seller performs calculations in ciphertext but lacks the decryption key). This design will effectively prevent the risk of sensitive data leakage from either party.

[0010] Furthermore, in order to prevent sellers from selling unreliable data or buyers from carefully designing calculation methods to extract the original data belonging to the data seller, based on the proposed bilateral secure transaction framework, the calculation methods of both parties can be adjusted according to the credibility of the participants (such as honest but curious, malicious and untrustworthy, etc.) to achieve secure calculation. Even if there are dishonest sellers or buyers, the data can only be processed in an encrypted state to avoid unauthorized data leakage. Ultimately, the buyer cannot grasp the original data, and it is difficult to use it for a new round of resale. This method not only provides a highly flexible privacy protection mechanism, but also ensures that the interests of all parties in the transaction process are protected.

[0011] Furthermore, the secure transaction can be privacy-preserving computing based on homomorphic encryption. For example, the specific implementation of linear regression training based on homomorphic encryption includes the following steps:

[0012] Step 1: System initialization phase: The data seller, as the data owner, holds the feature data set X and the target value Y. Assume that the number of rows and columns of the sample is I and J, respectively. The data buyer, as the model owner, holds the learning rate η and initializes the model weight W (a total of J dimensions: w1, w2, ..., w j ) and bias term b. In addition, the data buyer needs to specify important parameters such as the reduction function f and the training round epoch. To protect privacy, the data buyer initializes a fully homomorphic encryption key pair (pk, sk) and sends the public key pk to the data seller.

[0013] Specifically, the fully homomorphic encryption has additive homomorphism (denoted as in represents the addition operation on the ciphertext) and multiplication homomorphism (denoted by in It allows any number of addition and multiplication operations to be performed directly on the encrypted data.

[0014] Step 2, data encryption stage: the data buyer uses the public key pk to encrypt the data using the encryption algorithm Enc and in, The data buyer then sends the ciphertext Sent to data sellers.

[0015] Step 3, secure calculation phase: the data seller will use its data (X and Y) to complete the model according to the function f requirements ( and ) training task. The specific implementation of f is as follows:

[0016] ⑤The data seller calculates the ciphertext prediction value of each sample and ciphertext error value

[0017]

[0018] Where ∑ represents the accumulation on the ciphertext, this step can be accomplish.

[0019] ⑥ The data seller calculates the weight (where j∈[J]) and the bias term The partial derivative of

[0020]

[0021] Furthermore, we need to (where j∈[J]) and Complete the normalization process, that is, the goal is to get and

[0022] against The data seller and the data buyer collaborate in the calculation, and the calculation process is as follows:

[0023] (4) To protect privacy, the data seller generates a random number r1 and then calculates And call the encryption algorithm Enc to encrypt Then calculate send to said data buyer.

[0024] (5) The data buyer receives the data from the seller and uses the private key sk to call the decryption algorithm Dec to obtain

[0025] Then calculate the normalized mask result Call the encryption algorithm Enc to re-encrypt return To data sellers.

[0026] (6) The data seller calculates The normalized result can be obtained. The data seller and the data buyer collaborate in the calculation, and the calculation process is as follows:

[0027] (4) To protect privacy, the data seller generates a random number r j , then calculate And call the encryption algorithm Enc to encrypt Then calculate send to said data buyer.

[0028] (5) The data buyer receives the data from the seller and uses the private key sk to call the decryption algorithm Dec to obtain

[0029] Then calculate the normalized mask result Call the encryption algorithm Enc to re-encrypt return To data sellers.

[0030] (6) The data seller calculates The normalized result can be obtained.

[0031] ⑦ The data seller updates the model parameters in ciphertext

[0032]

[0033] ⑧If the training round epoch is reached, exit the loop and the final result is the updated (where j∈[J]) and

[0034] Otherwise, continue to repeat steps ① to ③.

[0035] Step 4: Result return stage: The data seller will train the final ciphertext model (where j∈[J]) and The result is returned to the data buyer for confirmation. The data buyer uses the private key sk to decrypt the final model through the decryption algorithm Dec, that is, calculate (where j∈[J]) and Thus, the final model of the plaintext is obtained, and the secure linear regression training task is completed.

[0036] In particular, in this model, privacy is protected by homomorphic encryption technology, and according to the designed steps, even if the data buyer initializes the model weights W to all 1 (total J dimensions: w1 = 1, w2 = 1, ..., w j =1) and the bias term b = 0, the calculation of wx+b will not leak privacy information. Because most operations on prediction value and error calculation are performed on the data seller side, data buyers can only coordinate passively, thus preventing malicious participants.

[0037] Furthermore, the secure transaction can also be based on privacy-preserving computing with secret sharing as the underlying layer. Under this model, it is assumed that the data seller and the data buyer are only interested in each other's data, but they fulfill the agreement in accordance with the law and will not intentionally manipulate values ​​or steal sensitive information. For example, the specific implementation of linear regression training based on secret sharing includes the following steps:

[0038] Step 1: System initialization phase: The data seller, as the data owner, holds the feature data set X and the target value Y. Assume that the number of rows and columns of the sample is I and J, respectively. The data buyer, as the model owner, holds the learning rate η and initializes the model weight W (a total of J dimensions: w1, w2, ..., w j ) and bias term b. In addition, the data buyer needs to specify important parameters such as the reduction function f and the training round epoch. To protect privacy, the data seller splits the feature data set X into two parts, namely X = X1 + X2, which means that for i∈[I], j∈[J], split x i，j =x i,j,1 +x i,j,2 , where X1 includes x i,j，1 part, X2 includes x i，j,2This process is not a simple sample splitting, but splitting each feature value into two parts, that is, secret sharing. Otherwise, if the sample size is simply split into two sub-sample sets, privacy will be leaked. Then X2 is sent to the data buyer. The data buyer only provides <f, epoch, W> to the data seller.

[0039] Step 2, secure calculation phase: The data seller and the data buyer each complete the operation according to function f, which is implemented as follows:

[0040] ⑤ The data seller and the data buyer each calculate the error value error for the X1 / X2 they hold (including the error value error of each sample i ), without leaking the predicted value and the original target value, the calculation process is:

[0041] (1) The data seller locally calculates Y-WX1 and sends Y-WX1 to the data buyer.

[0042] (2) The data buyer calculates WX2+b locally, and after receiving part of the data from the data seller, further calculates the error value error=Y-WX1-(WX2+b)

[0043] (3) The data buyer sends an error to the data seller for further processing.

[0044] ⑥Calculate weight Δw j (where j∈[J]) and the partial derivative of the bias term Δb, the calculation process is:

[0045] (1) The data seller uses the error value error to calculate the weighted partial derivative And sent to the data buyer.

[0046] (2) The data buyer uses the error value error to calculate the partial derivative of the bias term

[0047] (3) The data buyer normalization processing derivative: To get a more appropriate step size for gradient descent.

[0048] ⑦ The data buyer updates the model parameters locally

[0049]

[0050] ⑧If the training round epoch is reached, exit the loop, and the final result is the updated w i (where j∈[J]) and b, otherwise continue to repeat steps ① to ③.

[0051] Furthermore, as a technology that can perform calculations directly in the ciphertext state, the homomorphic encryption can effectively support the preprocessing link in the calculation process. Specifically, the complete calculation process can be divided into two stages: offline and online. In the offline stage, each participant generates a random number or constant, or uses homomorphic encryption technology to process these values, thereby refreshing the ciphertext. These random numbers or constants are stored in the form of one-time tuples and regenerated when they are insufficient; in the online stage, each party extracts the corresponding tuple as needed, and adds the random number therein as a mask to the input data, thereby ensuring that the privacy of the actual input data is protected during the calculation process. This phased approach will further improve the efficiency of online computing and protect data privacy, optimize the computational burden during protocol execution, and is particularly suitable for large-scale data transaction scenarios.

[0052] Furthermore, the secure two-party computation can further improve computational efficiency through parallel computing operations. Specifically, any participant can split a complex computational task into multiple independent subtasks for parallel execution (for example, computing a linear operation wx+b), and finally summarize the computational results to reduce the overall computational time and improve the efficiency of the protocol, thereby optimizing the utilization of computing resources.

[0053] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0054] 1. This invention provides a privacy-protected decentralized over-the-counter secure data trading method, which innovatively introduces secure computing methods between data sellers and buyers to balance data security and availability. Through two specific secure trading examples (① linear regression training based on homomorphic encryption, ② linear regression training based on secret sharing), it overcomes the limitations of the traditional centralized intermediary data hosting model and gives operational autonomy and data trading flexibility to data owners.

[0055] 2. The present invention can effectively solve the problems of data anti-copying and anti-secondary transactions, and transform these challenges into data processing problems. Based on the idea of ​​splitting the data set, or using homomorphic encryption to complete ciphertext operations, combined with secure multi-party computing, it ensures the confidentiality and integrity of the data of all parties in the transaction process, prevents untrusted participants from tampering with data, effectively responds to the risk of different trust levels among participants, and prevents malicious behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Schematic diagram of the system model of the method of the present invention. Specific implementation methods

[0058] In order to make the purpose, technical solution and advantages of the present invention more clear, the technical solution in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] like Figure 1 As shown, the present invention provides a privacy-preserving over-the-counter decentralized secure data trading method.

[0060] The secure transaction requires two-way security for data sellers and data buyers. Taking linear regression training based on homomorphic encryption as an example, the specific implementation of secure transactions includes the following steps:

[0061] Step 1: System initialization phase: The data seller, as the data owner, holds the feature data set X and the target value Y. Assume that the number of rows and columns of the sample is I and J, respectively. The data buyer, as the model owner, holds the learning rate η and initializes the model weight W (a total of J dimensions: w1, w2, ..., w j ) and bias term b. In addition, the data buyer needs to specify important parameters such as the reduction function f and the training round epoch. To protect privacy, the data buyer initializes a fully homomorphic encryption key pair (pk, sk) and sends the public key pk to the data seller.

[0062] Specifically, the fully homomorphic encryption has additive homomorphism (denoted as in represents the addition operation on the ciphertext) and multiplication homomorphism (denoted by in It allows any number of addition and multiplication operations to be performed directly on the encrypted data.

[0063] Step 2, data encryption stage: the data buyer uses the public key pk to encrypt the data using the encryption algorithm Enc and in, The data buyer then sends the ciphertext Sent to data sellers.

[0064] Step 3, secure calculation phase: the data seller will use its data (X and Y) to complete the model according to the function f requirements ( and ) training task. The specific implementation of f is as follows:

[0065] ① The data seller calculates the ciphertext prediction value of each sample and ciphertext error value

[0066]

[0067] Where ∑ represents the accumulation on the ciphertext, this step can be accomplish.

[0068] ② The data seller calculates the weight (where j∈[J]) and the bias term The partial derivative of

[0069]

[0070] Furthermore, we need to (where j∈[J]) and Complete the normalization process, that is, the goal is to get and

[0071] against The data seller and the data buyer collaborate in the calculation, and the calculation process is as follows:

[0072] (1) To protect privacy, the data seller generates a random number r1 and then calculates And call the encryption algorithm Enc to encrypt Then calculate send to said data buyer.

[0073] (2) The data buyer receives the data from the seller and uses the private key sk to call the decryption algorithm Dec to obtain

[0074] Then calculate the normalized mask result Call the encryption algorithm Enc to re-encrypt return To data sellers.

[0075] (3) The data seller calculates The normalized result can be obtained.

[0076] against The data seller and the data buyer collaborate in the calculation, and the calculation process is as follows:

[0077] (1) To protect privacy, the data seller generates a random number r j , then calculate And call the encryption algorithm Enc to encrypt Then calculate send to said data buyer.

[0078] (2) The data buyer receives the data from the seller and uses the private key sk to call the decryption algorithm Dec to obtain

[0079] Then calculate the normalized mask result Call the encryption algorithm Enc to re-encrypt return To data sellers.

[0080] (3) The data seller calculates The normalized result can be obtained.

[0081] ③The data seller updates the model parameters in ciphertext

[0082]

[0083] ④If the training round epoch is reached, exit the loop and the final result is the updated (where j∈[J]) and

[0084] Otherwise, continue to repeat steps ① to ③.

[0085] Step 4: Result return stage: The data seller will train the final ciphertext model (where j∈[J]) and The result is returned to the data buyer for confirmation. The data buyer uses the private key sk to decrypt the final model through the decryption algorithm Dec, that is, calculate (where j∈[J]) and Thus, the final model of the plaintext is obtained, and the secure linear regression training task is completed.

[0086] In particular, in this model, privacy is protected by homomorphic encryption technology, and according to the designed steps, even if the data buyer initializes the model weights W to all 1 (total J dimensions: w1 = 1, w2 = 1, ..., w j =1) and the bias term b = 0, the calculation of wx+b will not leak privacy information. Because most operations on prediction value and error calculation are performed on the data seller side, data buyers can only coordinate passively, thus preventing malicious participants.

[0087] On the other hand, taking linear regression training based on secret sharing as an example, the specific implementation of secure transactions includes the following steps: Step 1, system initialization phase: the data seller, as the data owner, holds the feature data set X and the target value Y. Assume that the number of rows and columns of the sample size is I and J; the data buyer, as the model owner, holds the learning rate η and initializes the model weights W (a total of J dimensions: w1, w2, ..., wj ) and bias term b. In addition, the data buyer needs to specify important parameters such as the reduction function f and the training round epoch. To protect privacy, the data seller splits the feature data set X into two parts, namely X = X1 + X2, which means that for i∈[I], j∈[J], split x i，j =x i，j,1 +x i，j，2 , where X1 includes x i，j，1 part, X2 includes x i，j，2 This process is not a simple sample splitting, but splitting each feature value into two parts, that is, secret sharing. Otherwise, if the sample size is simply split into two sub-sample sets, privacy will be leaked. Then X2 is sent to the data buyer. The data buyer only provides <f, epoch, W> to the data seller.

[0088] Step 2, secure calculation phase: The data seller and the data buyer each complete the operation according to function f, which is implemented as follows:

[0089] ① The data seller and the data buyer each calculate the error value error for the X1 / X2 they hold (including the error value error of each sample i ), without leaking the predicted value and the original target value, the calculation process is:

[0090] (1) The data seller locally calculates Y-WX1 and sends Y-WX1 to the data buyer.

[0091] (2) The data buyer calculates WX2+b locally, and after receiving part of the data from the data seller, further calculates the error value error=Y-WX1-(WX2+b)

[0092] (3) The data buyer sends an error to the data seller for further processing.

[0093] ② Calculate weight Δw j (where j∈[J]) and the partial derivative of the bias term Δb, the calculation process is:

[0094] (1) The data seller uses the error value error to calculate the weighted partial derivative And sent to the data buyer.

[0095] (2) The data buyer uses the error value error to calculate the partial derivative of the bias term

[0096] (3) The data buyer normalization processing derivative: To get a more appropriate step size for gradient descent.

[0097] ③The data buyer updates the model parameters locally

[0098]

[0099] ④ If the training round epoch is reached, it means that the linear regression training is finished and the loop is exited. The final result is the updated w j

[0100] (where j∈[J]) and b. Otherwise, continue to repeat steps ① to ③.

[0101] The above implementation examples are only used to illustrate the technical solution of the present invention rather than to limit it. A person skilled in the art may modify or convert the technical solution of the present invention without departing from the spirit and scope of the present invention. The scope of protection of the present invention shall be based on the claims.

Claims

1. A privacy-preserving over-the-counter decentralized secure data trading method, characterized in that: This method bypasses traditional intermediary brokers (i.e., third parties responsible for centralized data processing) and relies solely on data sellers and data buyers to complete transactions directly. As two independent computing entities, data sellers and data buyers will complete data processing under privacy protection by collaboratively executing two-party secure computing operations (e.g., based on homomorphic encryption, secret sharing, obfuscated circuits, etc.) such as model training, reasoning, and statistical analysis. Under this architecture, either the data seller or the data buyer holds private input (e.g., the data seller provides a data set, and the data buyer provides a training task or an initial model to be trained). Under the constraints of a given function (i.e., the prescribed operation steps), both parties are allowed to jointly calculate their respective private inputs and obtain the calculation results. In this process, it is ensured that neither party can know the other party's private input. This step can be formally expressed as: in is the encrypted data set provided by the seller, f is the established function provided by the buyer (the function is irreversible), and This is the return result obtained by the data buyer and does not include the original data X, thereby ensuring that only the right to use the data is traded, not the ownership of the data.

2. A privacy-protected over-the-counter decentralized secure data trading method according to claim 1, characterized in that: The secure transaction requires two-way security for data sellers and data buyers: on the one hand, data buyers can only obtain the results returned by the calculation, but cannot obtain the original data belonging to the data seller; on the other hand, data sellers cannot understand the intermediate calculation results and the private input provided by the data buyer (for example, the data seller performs calculations in ciphertext but lacks the decryption key). This design will effectively prevent the risk of sensitive data leakage from either party. Furthermore, in order to prevent sellers from selling unreliable data or buyers from carefully designing calculation methods to extract the original data belonging to the data seller, based on the proposed bilateral secure transaction framework, the calculation methods of both parties can be adjusted according to the credibility of the participants (such as honest but curious, malicious and untrustworthy, etc.) to achieve secure calculation. Even if there are dishonest sellers or buyers, the data can only be processed in an encrypted state to avoid unauthorized data leakage. Ultimately, the buyer cannot grasp the original data, and it is difficult to use it for a new round of resale. This method not only provides a highly flexible privacy protection mechanism, but also ensures that the interests of all parties in the transaction process are protected.

3. According to claim 1, a privacy-protected over-the-counter decentralized secure data trading method is characterized in that: The secure transaction can be privacy-preserving computing based on homomorphic encryption. For example, the specific implementation of linear regression training based on homomorphic encryption includes the following steps: Step 1: System initialization phase: The data seller, as the data owner, holds the feature data set X and the target value Y. Assume that the number of rows and columns of the sample is I and J, respectively. The data buyer, as the model owner, holds the learning rate η and initializes the model weight W (a total of J dimensions: w1, w2, ..., w j ) and bias term b. In addition, the data buyer needs to specify important parameters such as the reduction function f and the training round epoch. To protect privacy, the data buyer initializes a fully homomorphic encryption key pair (pk, sk) and sends the public key pk to the data seller. Specifically, the fully homomorphic encryption has additive homomorphism (denoted as in represents addition operation on ciphertext) and multiplication homomorphism (denoted as in It allows any number of addition and multiplication operations to be performed directly on the encrypted data. Step 2: Data encryption phase: The data buyer uses the public key pk to encrypt the data using the encryption algorithm Enc. and in, The data buyer then sends the ciphertext Sent to data sellers. Step 3, secure calculation phase: the data seller will use its data (X and Y) to complete the model according to the function f requirements ( and ) training task. The specific implementation of f is as follows: ① The data seller calculates the ciphertext prediction value of each sample and ciphertext error value Where ∑ represents the accumulation on the ciphertext, this step can be accomplish. ② The data seller calculates the weight (where j∈[J]) and the bias term The partial derivative of Furthermore, we need to (where j∈[J]) and Complete the normalization process, that is, the goal is to get and against The data seller and the data buyer collaborate in the calculation, and the calculation process is as follows: (1) To protect privacy, the data seller generates a random number r1 and then calculates And call the encryption algorithm Enc to encrypt Then calculate send to said data buyer. (2) The data buyer receives the data from the seller and uses the private key sk to call the decryption algorithm Dec to obtain Then calculate the normalized mask result Call the encryption algorithm Enc to re-encrypt return To data sellers. (3) The data seller calculates The normalized result can be obtained. The data seller and the data buyer collaborate in the calculation, and the calculation process is as follows: (1) To protect privacy, the data seller generates a random number r j , then calculate And call the encryption algorithm Enc to encrypt Then calculate send to said data buyer. (2) The data buyer receives the data from the seller and uses the private key sk to call the decryption algorithm Dec to obtain Then calculate the normalized mask result Call the encryption algorithm Enc to re-encrypt return To data sellers. (3) The data seller calculates The normalized result can be obtained. ③The data seller updates the model parameters in ciphertext ④If the training round epoch is reached, exit the loop and the final result is the updated (where j∈[J]) and Otherwise, continue to repeat steps ① to ③. Step 4: Result return stage: The data seller will train the final ciphertext model (where j∈[J]) and The result is returned to the data buyer for confirmation. The data buyer uses the private key sk to decrypt the final model through the decryption algorithm Dec, that is, calculate (where j∈[J]) and Thus, the final model of the plaintext is obtained, and the secure linear regression training task is completed. In particular, in this model, privacy is protected by homomorphic encryption technology, and according to the designed steps, even if the data buyer initializes the model weights W to all 1 (total J dimensions: w1 = 1, w2 = 1, ..., w j =1) and the bias term b = 0, the calculation of wx+b will not leak privacy information. Because most operations on prediction value and error calculation are performed on the data seller side, data buyers can only coordinate passively, thus preventing malicious participants.

4. According to claim 1, a privacy-protected over-the-counter decentralized secure data trading method is characterized in that: The secure transaction can also be based on privacy-preserving computing with secret sharing as the underlying layer. Under this model, it is assumed that the data seller and the data buyer are only interested in each other's data, but they fulfill the agreement in accordance with the law and will not intentionally manipulate values ​​or steal sensitive information. For example, the specific implementation of linear regression training based on secret sharing includes the following steps: Step 1: System initialization phase: The data seller, as the data owner, holds the feature data set X and the target value Y. Assume that the number of rows and columns of the sample is I and J, respectively. The data buyer, as the model owner, holds the learning rate η and initializes the model weight W (a total of J dimensions: w1, w2, ..., w j ) and bias term b. In addition, the data buyer needs to specify important parameters such as the reduction function f and the training round epoch. To protect privacy, the data seller splits the feature data set X into two parts, namely X = X1 + X2, which means that for i∈[I], j∈[J], split x i，j =x i，j，1 +x i,j,2 , where X1 includes x i，j,1 part, X2 includes x i,j,2 This process is not a simple sample splitting, but splitting each feature value into two parts, that is, secret sharing. Otherwise, if the sample size is simply split into two sub-sample sets, privacy will be leaked. Then X2 is sent to the data buyer. The data buyer only provides<f,epoch,W> To data sellers. Step 2, secure calculation phase: The data seller and the data buyer each complete the operation according to function f, which is implemented as follows: ① The data seller and the data buyer each calculate the error value error for the X1 / X2 they hold (including the error value error of each sample i ), without leaking the predicted value and the original target value, the calculation process is: (1) The data seller locally calculates Y-WX1 and sends Y-WX1 to the data buyer. (2) The data buyer calculates WX2+b locally, and after receiving part of the data from the data seller, further calculates the error value error=Y-WX1-(WX2+b) (3) The data buyer sends an error to the data seller for further processing. ② Calculate weight Δw j (where j∈[J]) and the partial derivative of the bias term Δb, the calculation process is: (1) The data seller uses the error value error to calculate the weighted partial derivative And sent to the data buyer. (2) The data buyer uses the error value error to calculate the partial derivative of the bias term (3) The data buyer normalization processing derivative: To get a more appropriate step size for gradient descent. ③The data buyer updates the model parameters locally ④ If the training round epoch is reached, exit the loop, and the final result is the updated w j (where j∈[J]) and b, otherwise continue to repeat steps ① to ③.

5. According to claim 1, a privacy-protected over-the-counter decentralized secure data trading method is characterized in that: As a technology that can perform calculations directly in the ciphertext state, homomorphic encryption can effectively support the preprocessing link in the calculation process. Specifically, the complete calculation process can be divided into two stages: offline and online. In the offline stage, each participant generates a random number or constant, or uses homomorphic encryption technology to process these values, thereby refreshing the ciphertext. These random numbers or constants are stored in the form of one-time tuples and regenerated when they are insufficient. At In the online phase, each party extracts the corresponding tuple as needed and adds the random number in it as a mask to the input data, thereby ensuring that the privacy of the actual input data is protected during the calculation process. This phased approach will further improve the efficiency of online computing and protect data privacy, optimize the computational burden during protocol execution, and is particularly suitable for large-scale data transaction scenarios. Furthermore, the secure two-party computation can further improve computational efficiency through parallel computing operations. Specifically, any participant can split a complex computational task into multiple independent subtasks for parallel execution (for example, computing a linear operation wx+b), and finally summarize the computational results to reduce the overall computational time, improve the efficiency of the protocol, and optimize the use of computing resources.

Images