Machine learning classification privacy protection method and device based on homomorphic encryption and storage medium

Abstract

The invention relates to a machine learning classification privacy protection method and device based on homomorphic encryption and a storage medium, and the method comprises the steps: supposing that a data owner does not have the homomorphic encryption and decryption capability, introducing a conversion mechanism of an AES ciphertext and a homomorphic ciphertext into a participant, and guaranteeing that the data owner does not have the homomorphic encryption and decryption capability except the data owner; other participants cannot pry the plaintext information of the data. According to the method, on the premise that the security is not reduced, the requirement on the computing power of a data owner is reduced, and the overall efficiency of the model is not excessively influenced.

Description

Technical Field

[0001] This patent application belongs to the field of machine learning algorithm technology, and more specifically, to a machine learning classification privacy protection method, device and storage medium based on homomorphic encryption. Background Art

[0002] With the rapid development of machine learning technology, more and more machine learning algorithms rely on cloud computing. Since cloud computing machine learning systems can obtain data from the user side when training or classifying in the cloud environment, concerns about data privacy and security are also increasing. Homomorphic encryption is a good solution, but it requires a lot of computing power. When users do not have homomorphic encryption capabilities and need to outsource data computing, the privacy protection issue after the data is out of the domain should be worthy of attention.

[0003] In the existing patent, "Machine Learning Linear Model Prediction Method Based on Homomorphic Encryption", announcement number: CN113704778A, the invention provides a machine learning linear model prediction method based on homomorphic encryption, including the following steps: the server side publishes the existing model and input and output field parameters; the client sends the model number to be used to the server side; the server side obtains the model weight parameters and training data preprocessing parameters according to the received number; the server side encrypts to obtain the encrypted model, and generates a pair of public keys and private keys at the same time, and sends the encrypted model and public key to the client; the client uses the public key and the encrypted model and the data to be predicted to perform linear operations to obtain an encrypted prediction result, and sends it to the server side; the server side receives the encrypted prediction result, uses the private key to decrypt the prediction result and returns it to the client, and the client obtains the prediction result.

[0004] The invention CN113704778A believes that both the client and the server have the ability of homomorphic encryption and decryption and homomorphic computing, which places certain high requirements on the computing power of the participants. However, in actual operation, the computing power of the data owner is often insufficient to complete homomorphic encryption and decryption, that is, the computing power requirements of the participants are too high, which will have too many negative effects on the overall efficiency of the model.

[0005] As can be seen from the above, the prior art has the following disadvantages:

[0006] 1. For general homomorphic encryption algorithms, due to their high mathematical complexity, their computational efficiency is often unbearable during the implementation process.

[0007] 2. Existing solutions only focus on linear machine learning models. For high-dimensional data, more complex machine learning models may be required.

[0008] 3. Since homomorphic encryption and decryption consume excessive computing resources, general data owners may not have the ability to directly perform homomorphic encryption on their data. Existing technologies do not consider the privacy protection issues after data leaves the domain when the data owner is unable to perform homomorphic encryption calculations.

[0009] As the informatization of the data industry has led to an explosive growth of data, the potential benefits of big data have brought great risks to data security. In order to solve the security issues of user privacy data in the era of rapid development of big data, a new machine learning solution is needed. Summary of the invention

[0010] The technical problem to be solved by the present invention is to overcome the above-mentioned shortcomings mentioned in the background and provide a machine learning classification privacy protection method, device and storage medium based on homomorphic encryption, which can not only protect the security of user privacy data, but also improve computing efficiency, shorten calculation time, and make the plaintext space better adapt to complex machine learning calculations.

[0011] In order to solve the above problems, the technical solution adopted by the present invention is:

[0012] The privacy protection method for machine learning classification based on homomorphic encryption is characterized by: assuming that the data owner does not have the ability of homomorphic encryption and decryption, by introducing the conversion mechanism of AES encryption and homomorphic encryption between the data owner and the model demander, it is ensured that except for the data owner, other participants cannot spy on the plaintext information of the data.

[0013] The specific steps include:

[0014] S1. The data owner encrypts the information to be classified using AES to form AES ciphertext and AES key, and sends the AES ciphertext to the model demander;

[0015] S2. The model demander homomorphically encrypts the AES ciphertext to form homomorphic ciphertext data, and uploads the homomorphic ciphertext data to the cloud server;

[0016] S3, the data owner sends the AES key to the cloud server;

[0017] S4. After receiving the homomorphic ciphertext data and AES key, the cloud server homomorphically decrypts the AES ciphertext, completes the homomorphic machine learning classification task according to the requirements of the model demander, and returns the homomorphic calculation result ciphertext to the model demander;

[0018] S5. The model demander homomorphically decrypts the data and obtains the required homomorphic computing results.

[0019] Furthermore, in S1, the data set into which the information to be classified is collected is an open source data set.

[0020] Furthermore, in S2, homomorphic encryption adopts the CKKS fully homomorphic encryption algorithm.

[0021] Furthermore, the efficiency of homomorphic multiplication in the CKKS fully homomorphic encryption algorithm is optimized.

[0022] Furthermore, "optimizing the efficiency of homomorphic multiplication in the CKKS fully homomorphic encryption algorithm" means performing rotation operations and Hadamard product operations on the AES encryption matrix according to the form of the AES encryption matrix, thereby greatly shortening the conversion time between AES ciphertext and homomorphic ciphertext data.

[0023] Further, "according to the form of the AES encryption matrix, a rotation operation and a Hadamard product operation are performed on the AES encryption matrix", specifically:

[0024] For a 4X4 original matrix A, expand it in rows

[0025]

[0026] In subsequent operations, use the Row Shifter operation multiple times to "cycle" the rows of this result "once";

[0027] In order to realize the multiplication operation of AES encryption matrix, the elements of each row of the original matrix A are summed

[0028]

[0029] With the above basic operators, we can combine the optimization ideas of matrix multiplication. Taking matrix A multiplied by matrix B as an example, the algorithm is as follows:

[0030]

[0031] Then apply RowShifter to matrix B and repeat the above operation. Repeat this operation for 4 rounds to get all the required target matrix values. Add the results homomorphically to get the homomorphic matrix multiplication result C.

[0032] Furthermore, the significant reduction means that the storage space and computational complexity of the ciphertext after homomorphic encryption are increased by more than a thousand times compared to the plaintext.

[0033] The scheme requires 4 calculations and 4 rotations between ciphertexts. For general matrix multiplication, 64 operations are required, and the matrix multiplication operation has been reduced to a polynomial level.

[0034] Furthermore, in S4, the application scenarios of homomorphic machine learning classification tasks include but are not limited to one or more of SVM, K-means and neural networks.

[0035] A machine learning classification privacy protection device based on homomorphic encryption, used to implement the above method, including

[0036] AES encryption module, used by the data owner to encrypt the information to be classified by AES, form AES ciphertext and AES key, and send the AES ciphertext to the model demander;

[0037] Homomorphic encryption module, used by the model demander to perform homomorphic encryption on the AES ciphertext to form homomorphic ciphertext data, and upload the homomorphic ciphertext data to the cloud server;

[0038] AES key sending module, used by the data owner to send the AES key to the cloud server;

[0039] The cloud computing module is used for the cloud server to homomorphically decrypt the AES ciphertext after receiving the homomorphic ciphertext data and AES key, complete the homomorphic machine learning classification task according to the requirements of the model demander, and return the homomorphic calculation result ciphertext to the model demander;

[0040] The homomorphic decryption module is used by the model demander to homomorphically decrypt data and obtain the required homomorphic computing results.

[0041] A computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method of claims 1-8.

[0042] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are:

[0043] 1. AES ciphertext conversion and machine learning classification algorithm calculations under homomorphic encryption involve a large number of homomorphic addition and homomorphic multiplication calculations. Among them, CKKS homomorphic multiplication calculations also involve rotation operations on the ciphertext matrix. The time consumed by a rotation operation is about twenty times that of a homomorphic addition, and the time consumed by a homomorphic multiplication is about forty times that of a homomorphic addition. Therefore, in order to improve the efficiency of the ciphertext conversion process and homomorphic encryption machine learning, the focus is on optimizing the efficiency of homomorphic multiplication in the process.

[0044] 2. A mechanism for converting AES to homomorphic ciphertext is proposed. An alternative solution is proposed for non-polynomial functions in the AES encryption and decryption process to improve the security solution in the data flow process and ensure the safe exit of user data.

[0045] 3. Through the cloud platform, some machine learning classification schemes were implemented, including SVM, K-Means scheme and neural network.

[0046] 4. In the present invention, except for the data owner, all participants cannot obtain the plain text information of the data, which effectively prevents other participants from illegally snooping on the information.

[0047] 5. Compared with the existing inventions, the present invention believes that the computing power of the data owner is not sufficient to complete homomorphic encryption and decryption. Compared with the inventions in the existing background, the present invention reduces the computing power requirements of the data owner without reducing security, and will not have too much impact on the overall efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a processing flow chart of the method of the present invention.

[0049] Figure 2 It is a connection block diagram of the device of the present invention. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below with reference to the embodiments.

[0051] Before introducing the present invention, it is necessary to introduce related concepts.

[0052] 1.AES encryption:

[0053] Advanced Encryption Standard (AES) is the most common symmetric encryption algorithm, and WeChat Mini Program uses this encryption algorithm for encrypted transmission. Symmetric encryption algorithm means that the same key is used for encryption and decryption. In the process of users transmitting their own medical data to the server, in order to prevent illegal eavesdroppers in the communication link, users will encrypt the data to be sent before uploading it. As an encryption method certified by the US federal government, AES encryption has the characteristics of simple encryption algorithm and high security. Therefore, we use AES encryption for ciphertext transmission. The AES encryption process involves 4 operations, namely byte substitution, row shift, column confusion and round key addition. The decryption process is the corresponding inverse operation. Since each step of the operation is reversible, the plaintext can be restored by decrypting in the opposite order. The key for each round in encryption and decryption is obtained by expanding the initial key. In the algorithm, the 16-byte plaintext, ciphertext and round key are all represented by a 4x4 matrix.

[0054] Since the plaintext, ciphertext and round key of the AES algorithm are all 4x4 matrices, there is no ciphertext expansion during the encryption process, and the security factor is higher because it is easier to perform decryption operations on it in the same way.

[0055] 2. Homomorphic encryption:

[0056] In order to achieve the complete conversion between AES and homomorphic encryption, the plaintext space of the data is expanded to the domain R with a floating point precision of 2-4 212We choose the current mainstream homomorphic encryption algorithm, CKKS, as the encryption algorithm for the homomorphic computing part of this model. The CKKS fully homomorphic encryption algorithm mainly consists of 7 parts: initialization (setup), key generation (keygeneration), encryption (encryption), decryption (decryption), addition (addition), multiplication (multiplication), and rescaling (rescaling). The above content is the existing technology.

[0057] (1) Initialization: Define parameter λ as the safety level, L as the upper limit of depth. Select a positive integer N that is approximately equal to a power of 2, a base p>0, p0>0, and a special modulus P>0 (for rescaling). Define Q = p0·p L , and N and P·Q satisfy the security level λ. It should be noted here that CKKS is a hierarchical fully homomorphic encryption, each ciphertext corresponds to a depth, and there are L layers in total. The modulus in the top layer is Q = p0·p L , followed by p0·p L-1 , p0·p L-2 , ..., p0. As the number of layers decreases, the modulus of the layer also decreases. This setting is to solve the problem of increasing ciphertext in multiplication. Once the ciphertext is large enough, the computational efficiency will decrease. Let q represent the modulus of any layer, where q is a divisor of Q. Choose a distribution χ related to the private key s , an error distribution χ e and a random distribution χ r .

[0058] (2) Key generation: represents a polynomial quotient ring, represents a polynomial quotient ring with Q as the modulus, randomly selecting s←χ s , and e←χ e . Set the private key sk←(1,s) and calculate the public key Here b = -a·s + emodQ. Random selection and e'←χ e , set the auxiliary calculation key Here b'=-a'·s+e'+P·s 2 modP·Q.

[0059] (3) Encryption: Randomly select r←χ r and e0,e1←χ e For the plaintext polynomial Output a ciphertext c←r·pk+(m+e0,e1)modQ.

[0060] (4) Decryption: For the ciphertext and the private key sk, calculate <c, sk> mod q = c0 + c1 mod q.

[0061] (5) Addition: For two ciphertexts Output c add ← c + c' mod q,

[0062] (6) Multiplication: Multiplication is divided into multiplying ciphertext by plaintext and multiplying two ciphertexts. For the former, given the ciphertext plaintext Output c mult ← m · c mod q, For the latter, given two ciphertexts Output

[0063] (7) Rescaling: Given the ciphertext and a new modulus q' < q, output where denotes the floor of (q' / q) · c.

[0064] Next, introduce the present invention: A privacy protection method for machine learning classification based on homomorphic encryption. Assume that the data owner does not have the ability of homomorphic encryption and decryption. By introducing a conversion mechanism of AES encryption and homomorphic encryption between the data owner and the model requester, it is ensured that except for the data owner, other participating parties cannot窥探 the plaintext information of the data.

[0065] The specific steps are as Figure 1 :

[0066] ① The data owner encrypts the information to be classified using AES to form an AES ciphertext and an AES key; the data set to which the information to be classified belongs is an open-source data set;

[0067] ② The data owner sends the AES ciphertext to the model requester;

[0068] ③ The model requester performs homomorphic encryption on the AES ciphertext to form homomorphic ciphertext data. The CKKS fully homomorphic encryption algorithm is used for homomorphic encryption, and the efficiency of homomorphic multiplication in the CKKS fully homomorphic encryption algorithm is optimized;

[0069] ④ The model requester uploads the homomorphic ciphertext data and evk data to the cloud server;

[0070] ⑤ The data owner sends the AES key to the cloud server;

[0071] ⑥ After receiving the homomorphic ciphertext data and the AES key, the cloud server performs homomorphic decryption on the AES ciphertext; It should be noted that the term "窥探" in the original text is not a common English expression in this context. A more appropriate word might be "peek at" or "snoop on". However, according to the requirements, the translation is kept as it is.

[0072] ⑦ The cloud server completes the homomorphic machine learning classification task according to the requirements of the model demander;

[0073] ⑧The cloud server returns the encrypted result of the homomorphic computing to the model demander;

[0074] ⑨The model demander homomorphically decrypts the data and obtains the required homomorphic computing results.

[0075] In the above process, ①-⑥ are transformation processes, ⑦-⑨ are calculation processes, which can completely correspond to S1-S5 of the claims.

[0076] It can be seen that this paper proposes a data privacy protection machine learning scheme based on homomorphic encryption and conducts experimental verification on it. The scheme is roughly divided into three parts in the homomorphic encryption machine learning scenario. In the first part, the data owner encrypts the information to be classified by AES and sends it to the model demander; in the second stage, the model demander homomorphically encrypts the AES ciphertext; in the third stage, the cloud server homomorphically decrypts the AES ciphertext, performs the machine learning classification task, and returns the calculation result to the user. It is worth noting that the present invention assumes that the data owner does not have the ability of homomorphic encryption. By introducing the conversion mechanism between AES ciphertext and homomorphic ciphertext, it ensures that except for the data owner, other participants cannot spy on the plaintext information of the data.

[0077] Meeting the requirements of machine learning

[0078] In our solution, AES ciphertext conversion and machine learning classification algorithm calculation under homomorphic encryption involve a large number of homomorphic addition and homomorphic multiplication calculations. Among them, CKKS homomorphic multiplication calculation also involves the rotation operation of the ciphertext matrix. The time consumed by a rotation operation is about 20 times that of homomorphic addition, and the time consumed by a homomorphic multiplication is about 40 times that of homomorphic addition. Therefore, in order to improve the efficiency of the ciphertext conversion process and homomorphic encryption machine learning, the focus is on optimizing the efficiency of homomorphic multiplication in the process. This paper proposes a solution that can pack more elements into a ciphertext.

[0079] For a 4X4 AES matrix, expand it in rows

[0080]

[0081] In subsequent operations, we need to use the Row Shifter operation multiple times to "cycle" the rows of this result "one round" (4 times).

[0082] To perform matrix multiplication, we need to sum the elements of each row of the original matrix.

[0083]

[0084] With the above basic operators, we can combine a set of optimization ideas for matrix multiplication. Taking matrix A multiplied by matrix B as an example, the algorithm is as follows:

[0085]

[0086] Then apply RowShifter to matrix B and repeat the above operation. Repeat this operation for 4 rounds and we can get all the target matrix values ​​we need. Add the result homomorphically and we get the homomorphic matrix multiplication result C. This scheme calculates 4 times and rotates 4 times between ciphertexts. For general matrix multiplication, 64 operations are required, and the matrix multiplication operation is reduced by polynomial level.

[0087] Machine learning classification privacy protection scheme based on homomorphic encryption:

[0088] This solution mainly takes into account that the model demander needs the sensitive data of the data owner to perform the required machine learning classification tasks. The computing power of the data owner is weak and is not enough to complete the task of homomorphically encrypting the data and then leaving the domain. Therefore, the present invention proposes a machine learning classification privacy protection solution based on homomorphic encryption. The model includes three participants: the data owner, the model demander, and the cloud platform. The computing power of the three participants is from large to small: cloud server, model demander, and data owner. We assume that the data owner can implement AES data encryption and realize the safe exit of its own sensitive data for the model demander to complete the required computing tasks. The model demander can calculate the homomorphic encryption and decryption of the data. The main task is to send a data demand request to the data owner and a data calculation request to the cloud platform to complete the computing tasks required by its own model and perform homomorphic encryption and decryption on the sensitive data of the data owner.

[0089] The cloud platform has all the computing power required by the model. Its main task is to complete the homomorphic machine learning classification task according to the requirements of the model demander. After the model demander completes the conversion between AES ciphertext and homomorphic ciphertext, different types of data sets are used to analyze the time efficiency and accuracy of SVM, K-means and neural network under homomorphic encryption. The experimental results show that the proposed solution is acceptable in terms of time efficiency and accuracy. This solution provides a safer idea for data flow and is more suitable for large-scale machine learning calculations.

[0090] This invention combines the currently popular and secure technologies, including AES encryption technology, homomorphic encryption technology, secret sharing technology, etc. to protect the transmission security and computing security of data throughout the life cycle. The key points of this technology are:

[0091] 1. Compared with plaintext, the storage space and computational complexity of the ciphertext after homomorphic encryption are improved by at least one thousand times. Taking into account that the computing power of the data owner is weak and the homomorphic encryption technology cannot be used directly, the present invention provides a new multi-party secure computing model involving three parties, which simplifies the computing power of the data owner, is more practical, and has guaranteed security. It ensures that the data owner's information cannot be snooped by other participants except the data owner itself, but the calculation results required by the model demander can still be completed, ensuring the calculation security. The homomorphic encryption technology used supports multiple addition, subtraction, multiplication and division operations compared with other homomorphic encryption algorithms, with high efficiency, and can complete the required homomorphic machine learning classification calculations more quickly.

[0092] 2. Based on key point 1, the present invention constructs an AES and homomorphic encryption conversion mechanism between the data owner and the model demander, ensuring that all participants except the data owner cannot obtain private data, thereby ensuring the security of private data.

[0093] 3. According to the form of the AES encryption matrix, the present invention performs rotation operations and Hadamard product operations on the encryption matrix to optimize the time complexity of matrix multiplication under homomorphic encryption, thereby greatly shortening the conversion time between AES ciphertext and homomorphic encryption ciphertext.

[0094] 4. The present invention is not limited to the AES encryption matrix. 28 The plaintext space on the finite field increases the input space of encrypted plaintext data, which is more conducive to subsequent homomorphic machine learning operations.

[0095] 5. The present invention uses the currently more efficient CKKS homomorphic encryption algorithm, which effectively improves the efficiency of homomorphic computing and shortens the overall computing time.

[0096] 6. Through the cloud platform, the homomorphic encryption scheme provided by the present invention is applied to several typical machine learning classification scenarios using open source data sets, including SVM, K-Means and neural networks. Experiments have shown that the scheme has certain advantages in terms of efficiency and security.

[0097] Another embodiment of the present invention provides a machine learning classification privacy protection device based on homomorphic encryption, such as Figure 2 , the device comprises

[0098] AES encryption module, used by the data owner to encrypt the information to be classified by AES, form AES ciphertext and AES key, and send the AES ciphertext to the model demander;

[0099] Homomorphic encryption module, used by the model demander to perform homomorphic encryption on the AES ciphertext to form homomorphic ciphertext data, and upload the homomorphic ciphertext data to the cloud server;

[0100] AES key sending module, used by the data owner to send the AES key to the cloud server;

[0101] The cloud computing module is not only responsible for receiving information, but also for the subsequent homomorphic machine learning computing tasks. Specifically, after the cloud server receives the homomorphic ciphertext data and AES key, it homomorphically decrypts the AES ciphertext, completes the homomorphic machine learning classification task according to the requirements of the model demander, and returns the homomorphic computing result ciphertext to the model demander;

[0102] The homomorphic decryption module is used by the model demander to homomorphically decrypt data and obtain the required homomorphic computing results.

[0103] The embodiment of the present invention also provides another computer-readable storage medium, which can be the computer-readable storage medium contained in the memory in the above embodiment; or it can be a computer-readable storage medium that exists independently and is not installed in the terminal. The computer-readable storage medium stores one or more programs, and the one or more programs are used by one or more processors to execute Figure 1 The illustrated embodiment provides a privacy protection method for machine learning classification based on homomorphic encryption.

Claims

1. A privacy protection method for machine learning classification based on homomorphic encryption, characterized by: Assuming that the data owner does not have the ability to encrypt and decrypt in a homomorphic manner, by introducing the AES encryption and homomorphic encryption conversion mechanism between the data owner and the model demander, it is ensured that other participants except the data owner cannot spy on the plaintext information of the data. The specific steps include: S1. The data owner encrypts the information to be classified using AES to form AES ciphertext and AES key, and sends the AES ciphertext to the model demander; S2. The model demander homomorphically encrypts the AES ciphertext to form homomorphic ciphertext data, and uploads the homomorphic ciphertext data to the cloud server; S3, the data owner sends the AES key to the cloud server; S4. After receiving the homomorphic ciphertext data and AES key, the cloud server homomorphically decrypts the AES ciphertext, completes the homomorphic machine learning classification task according to the requirements of the model demander, and returns the homomorphic calculation result ciphertext to the model demander; S5. The model demander homomorphically decrypts the data and obtains the required homomorphic computing results.

2. The privacy protection method for machine learning classification based on homomorphic encryption according to claim 1 is characterized in that: In S1, the data set that contains the information to be classified is an open source data set.

3. The privacy protection method for machine learning classification based on homomorphic encryption according to claim 1 is characterized in that: In S2, homomorphic encryption uses the CKKS fully homomorphic encryption algorithm.

4. The privacy protection method for machine learning classification based on homomorphic encryption according to claim 3 is characterized in that: Optimize the efficiency of homomorphic multiplication in the CKKS fully homomorphic encryption algorithm.

5. The privacy protection method for machine learning classification based on homomorphic encryption according to claim 4 is characterized in that: "Optimizing the efficiency of homomorphic multiplication in the CKKS fully homomorphic encryption algorithm" means that according to the form of the AES encryption matrix, the AES encryption matrix is ​​rotated and the Hadamard product is calculated, thereby greatly shortening the conversion time between AES ciphertext and homomorphic ciphertext data.

6. The privacy protection method for machine learning classification based on homomorphic encryption according to claim 5 is characterized in that: "According to the form of the AES encryption matrix, the AES encryption matrix is ​​rotated and Hadamard product is calculated", specifically: For a 4X4 original matrix A, expand it in rows In order to realize the multiplication operation of AES encryption matrix, the elements of each row of the original matrix A are summed With the above basic operators, we can combine the optimization ideas of matrix multiplication. Taking matrix A multiplied by matrix B as an example, the algorithm is as follows: Then apply RowShifter to matrix B and repeat the above operation. Repeat this operation for 4 rounds to get all the required target matrix values. Add the results homomorphically to get the homomorphic matrix multiplication result C.

7. The privacy protection method for machine learning classification based on homomorphic encryption according to claim 5 is characterized in that: The significant reduction means that the storage space and computational complexity of the ciphertext after homomorphic encryption are increased by more than a thousand times compared to the plaintext.

8. The privacy protection method for machine learning classification based on homomorphic encryption according to claim 1, characterized in that: In S4, the application scenarios of homomorphic machine learning classification tasks include but are not limited to one or more of SVM, K-means and neural network.

9. A privacy protection device for machine learning classification based on homomorphic encryption, characterized in that: The method for implementing any one of claims 1 to 8 above comprises: AES encryption module, used by the data owner to encrypt the information to be classified by AES, form AES ciphertext and AES key, and send the AES ciphertext to the model demander; Homomorphic encryption module, used by the model demander to perform homomorphic encryption on the AES ciphertext to form homomorphic ciphertext data, and upload the homomorphic ciphertext data to the cloud server; AES key sending module, used by the data owner to send the AES key to the cloud server; The cloud computing module is used for the cloud server to homomorphically decrypt the AES ciphertext after receiving the homomorphic ciphertext data and AES key, complete the homomorphic machine learning classification task according to the requirements of the model demander, and return the homomorphic calculation result ciphertext to the model demander; The homomorphic decryption module is used by the model demander to homomorphically decrypt data and obtain the required homomorphic computing results.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to claims 1 to 8 are implemented.

Images