CN113938266B - Junk mail filter training method and system based on integer vector homomorphic encryption - Google Patents

Abstract

The invention provides a spam filter training method and system based on integer vector homomorphic encryption, which includes the following steps: S1: preprocessing, including system parameters and key generation; S2: encrypting the training data set; S3: Privacy-protecting spam filtering model training; S4: The cloud server-side spam filtering model training is completed and the model weights are returned. The present invention can greatly ensure the security and reliability of data on the premise of realizing basic needs, and has high practical value and expansion space. In the model training protocol, most of the calculations are performed by the cloud server, so the client's workload is very low. When the client's computing power is limited, calculations can be made conveniently and quickly. In the client-based protocol, the client performs most of the calculations. Part of the calculation only occurs once, which also reduces communication overhead to a certain extent.

Description

Spam filter training method and system based on integer vector homomorphic encryption

Technical field

The present invention relates to the technical field of secure multi-party computing privacy protection, and specifically to a spam filter training method and system based on integer vector homomorphic encryption.

Background technique

In recent years, with the continuous development of machine learning technology, people have paid more and more attention to privacy protection. The rise of cloud computing has led to the development of "machine learning as a service", where enterprise users can outsource work to third-party providers. At the same time, cloud computing cannot violate users' data privacy. One way to protect privacy and security is to encrypt data before uploading it to the cloud. If encryption prevents efficient computation, the utility of the data will be severely limited. However, this problem can be solved by using a homomorphic encryption scheme that allows operations to be performed on encrypted data without decryption, without revealing the results. any information other than

Email is a private communication medium in which messages are read only by the recipient. Due to the sensitive nature of the information content, there may be personal, strategic and legal restrictions on sharing and publishing email data. These limitations create a significant barrier in many email processing applications, such as spam filtering, which are often provided by a separate service provider. The inherent privacy constraints of email are a major obstacle to developing effective spam filtering methods, which require access to large amounts of email data belonging to multiple users. To alleviate this problem, a privacy-preserving spam filtering system is designed in which the server is able to train and evaluate a spam classifier based on logistic regression, and uses techniques such as homomorphic encryption and randomization to make it impossible for all users who provide email data to Observed any data other than itself.

There are currently some methods to solve the above problems, such as data aggregation methods based on homomorphic encryption algorithms; simply put, the function of data aggregation is to aggregate multiple data into one data; and the homomorphic encryption algorithm has the following characteristics: Performing ring addition and multiplication operations on the plaintext and then encrypting it, and performing corresponding operations on the ciphertext after encryption are equivalent. Existing aggregation schemes use a paillier encryption scheme to complete model training, allowing the paillier to be used in the real number domain by expanding the data by a multiple of a large integer. However, such aggregation scheme has shortcomings such as low efficiency, cumbersome scheme, and shortened encryption domain.

Contents of the invention

The problem to be solved by the present invention is that existing spam filter training methods have problems such as poor performance and cumbersome steps.

In order to solve the above problems, on the one hand, the present invention provides a spam filter training method based on integer vector homomorphic encryption, which includes the following steps:

S1: Preprocessing, including system parameters and key generation;

S2: Encrypt the training data set;

S3: Privacy-preserving spam filtering model training;

S4: The cloud server-side spam filtering model training is completed, and the model weights are returned.

Preferably, step S1 specifically includes the following steps:

S11: Client data processing: First, the corresponding word vector is generated according to the content of the text data. The word vector is a vector in which words or phrases from the vocabulary are mapped to real numbers. For a processed text data, if the word exists in the sentence Then 1 is given, or 0 if it does not exist, and then the encryption key pair is generated;

S12: Cloud server-side data processing: First, some vectors and matrices are preset on the server for storage, including unit vectors and encoding vectors of various lengths.

Preferably, the step S2 specifically includes the following steps:

S21: The client uses the encryption function to encrypt the vector, uses the encryption tool to encrypt the original data set one by one, and then sends the encrypted data set, encryption tool function, public key and auxiliary information to the server;

S22: The client uploads the encrypted ciphertext data set and public key M to the cloud server. The cloud server receives the encrypted data set and then runs the homomorphic encryption logistic regression algorithm.

Preferably, the step S3 specifically includes the following steps:

S31: The cloud server sets the number of iterations n;

S32: Calculate the ciphertext gradient;

S33: Update parameters using gradient descent method;

S34: Repeat steps S32 and S33 until the iteration number n is reached.

Preferably, the step S4 specifically includes the following steps:

S41: After the training is completed, the cloud server obtains the encrypted training model and returns the ciphertext model and model weight ω to the client;

S42: The client receives the ciphertext model and model weight ω, and uses the private key to decrypt it to obtain the plaintext model, which is the trained spam filtering model, and uses the model to predict the data.

On the other hand, the present invention also provides a system that adopts the above-mentioned spam filter training method based on integer vector homomorphic encryption, as shown in Figure 2, wherein the system includes:

Client and cloud server;

The client is composed of a client host group and is used to provide a training data set and encrypt the training data set;

The cloud server is used to complete the encrypted data set spam filtering training model;

After the client uses the public key to encrypt the training data set, the training data set is sent to the cloud server. The cloud server runs the privacy protection logistic regression algorithm training model. After the training is completed, it is sent to the cloud server. The client uses the private key to decrypt to obtain the plaintext model.

Compared with the existing technology, the spam filter training method and system based on integer vector homomorphic encryption described in the present invention has the following beneficial effects:

(1) The spam filter training method and system based on integer vector homomorphic encryption according to the present invention includes a cloud server-based privacy protection logistic regression training protocol and a client-based privacy data encryption protocol. During the execution of the protocol, use Homomorphic encryption encrypts the client's private data and uploads it to the cloud server. The cloud server then uses these encrypted data to run the privacy-protecting logistic regression protocol. After the training is completed, the ciphertext of the model weights is obtained, and finally the cloud server Transmitted to the client, the client uses its private key to decrypt to obtain the plaintext model weight;

(2) The spam filter training method and system based on integer vector homomorphic encryption of the present invention uses the least squares method to fit the logistic function in the logistic regression training method, so that the final calculation result can be better fitted Data can greatly ensure the security and reliability of data while meeting basic needs, and has high practical value and room for expansion. In the model training protocol, most of the calculations are performed by the cloud server, so the client's workload is very low. When the client's computing power is limited, calculations can be made conveniently and quickly; in the client-based protocol, the client performs most of the calculations. Part of the calculation only occurs once, which also reduces communication overhead to a certain extent.

Description of the drawings

Figure 1 is a flow chart of the method of the present invention;

Figure 2 is a schematic diagram of the system of the present invention.

Detailed ways

In order to make the above objects, features and advantages of the present invention more obvious and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Embodiment 1

A spam filter training method based on integer vector homomorphic encryption is provided, as shown in Figure 1, which includes the following steps:

S1: Preprocessing, including system parameters and key generation;

S2: Encrypt the training data set;

S3: Privacy-preserving spam filtering model training;

S4: The cloud server-side spam filtering model training is completed, and the model weights are returned.

Among them, the step S1 specifically includes the following steps:

S11: Client data processing: First, the corresponding word vector is generated according to the content of the text data. The word vector is a vector in which words or phrases from the vocabulary are mapped to real numbers. For a processed text data, if the word exists in the sentence Then give 1, if it does not exist, give 0, and then generate the encryption key pair; specifically:

First, the client initializes the parameters and converts the email data set into a word vector data set; then constructs the private key S, S = [I, T], where I∈Z ^m×m is the identity matrix, and T∈Z ^m×(nm) is a random matrix. Then use the key conversion technology, first convert the key S into a new key S′ through bit conversion, which satisfies: S′c ^* = Sc, where c is the ciphertext and c ^* is the bit-converted ciphertext. ; Then establish a key conversion matrix M, which satisfies: S'M=S ^* +E, where E is the error term, and M is used as the public key for encrypted data.

S12: Cloud server-side data processing: First, some vectors and matrices are preset on the server for storage, including unit vectors and encoding vectors of various lengths, thereby reducing the calculation amount of the algorithm itself, thereby making the integer vector homomorphic encryption algorithm More efficient in ciphertext logistic regression schemes.

Among them, the step S2 specifically includes the following steps:

S21: The client uses the encryption function to encrypt the vector, uses the encryption tool to encrypt the original data set one by one, and then sends the encrypted data set, encryption tool function, public key and auxiliary information to the server; specifically:

For the word vector data set D={(x ₁ , y ₁ ), (x ₂ , y ₂ ),..., (x _m , y _m )} generated in step S11, use the integer vector encryption scheme to encrypt with the public key M as D={(c ₁ ,y ₁ ), (c ₂ ,y ₂ ),…,(c _m ,y _m )}, where c _i satisfies Sc _i =wx _i +e _i , where w is a large integer, e _i is a random error less than w/2.

S22: The client uploads the encrypted ciphertext data set and public key M to the cloud server. The cloud server receives the encrypted data set and then runs the homomorphic encryption logistic regression algorithm.

Among them, the step S3 specifically includes the following steps:

S31: The cloud server sets the number of iterations n; specifically:

The cloud server generates a random vector ρ and the number of iterations n.

S32: Calculate the ciphertext gradient; specifically:

Cloud server-side calculation of ciphertext gradient where h _ρ is the Sigmoid function,/> This function is approximated using the least squares method, i.e. calculated using the following formula:

g ₇ (x)＝b ₀ +b ₁ (x/8)+b ₃ (x/8) ³ +b ₅ (x/8) ⁵ +b ₇ (x/8) ⁷

Finally, all operations can be converted into addition and inner product operations between vectors.

S33: Use gradient descent method to update parameters; specifically:

Update parameters using gradient descent: where ρ _j is the iterative value of the jth training parameter.

S34: Repeat steps S32 and S33 until the iteration number n is reached.

Among them, the step S4 specifically includes the following steps:

S41: After the training is completed, the cloud server obtains the encrypted training model and returns the ciphertext model and model weight ω to the client;

S42: The client receives the ciphertext model and model weight ω, and uses the private key to decrypt it to obtain the plaintext model, which is the trained spam filtering model. This model can be used to predict the data.

In this way, the method in this embodiment implements the training of a privacy-preserving spam filter based on the integer vector homomorphic encryption algorithm and the privacy-preserving logistic regression algorithm. First, the client performs data processing: generates corresponding word vectors based on the content of the text data; then, runs an encryption algorithm to encrypt it to obtain ciphertext data and sends it to the cloud server; after the cloud server receives the ciphertext, it runs The privacy-preserving logistic regression algorithm performs model training under ciphertext; after the training is completed, the ciphertext model is obtained and sent back to the client. The client uses its private key to decrypt and obtain the plaintext model. This method can not only protect the privacy of training data, but also prevent the cloud server from obtaining training model information.

Embodiment 2

A system is provided, which adopts the spam filter training method based on integer vector homomorphic encryption as described in Embodiment 1, as shown in Figure 2, wherein the system includes:

Client (Client) and cloud server (Cloud);

The client is composed of a client host group, and is used to provide a training data set and encrypt the training data set;

The cloud server is used to complete the encrypted data set spam filtering training model;

After the client uses the public key to encrypt the training data set, the training data set is sent to the cloud server. The cloud server runs the privacy protection logistic regression algorithm training model. After the training is completed, it is sent to the cloud server. The client uses the private key to decrypt to obtain the plaintext model.

In this way, the system in this embodiment has two entities: the client and the cloud server; the client (Client) owns the email data set used to train the model. If it is not encrypted, the cloud server (Cloud) can directly obtain it. These data sets may contain some private and sensitive data, so the client does not want the cloud server to obtain the information. Therefore, the client first generates a key, uses the generated key to encrypt its own data set, and finally sends the encrypted data set, that is, the ciphertext data to the cloud server; the cloud server gets the ciphertext data and only It can be analyzed and processed according to the inherent computing protocol, so it will complete the ciphertext logistic regression training according to the preset ciphertext data training method, and finally a ciphertext logistic regression model will be obtained. At the same time, the cloud server cannot decrypt the model, so It also cannot obtain model information; finally, the cloud server sends the ciphertext model to the client.

Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims (4)

1. A training method of a spam filter based on integer vector homomorphic encryption is characterized by comprising the following steps:

s1: preprocessing, including system parameters and key generation;

the step S1 comprises the following steps:

s11: client data processing: firstly, generating a corresponding word vector according to the content of text data, wherein the word vector is a vector in which words or phrases from a vocabulary are mapped to real numbers, and for one processed text data, giving 1 if the vocabulary exists in a sentence, giving 0 if the vocabulary does not exist, and then regenerating an encryption key pair;

the client initializes parameters and converts the mail data set into a word vector data set; building a private key S, s= [ I, T]Wherein I is E Z ^m×m Is a unit matrix, T is E Z ^m×(n-m) The key S is converted into a new key S' by bit conversion using a key conversion technique for a random matrix, which satisfies the following conditions: s' c ^* Sc, where c is ciphertext, c ^* A ciphertext after bit conversion; establishing a key conversion matrix M which satisfies the following conditions: s' m=s ^* +E, where E is the error term, M being the public key of the encrypted data

S12: cloud server side data processing: firstly, presetting a plurality of vectors and matrixes on a server for storage, wherein the vectors and matrixes comprise unit vectors and coding vectors with various lengths;

s2: encrypting the training data set;

the step S2 comprises the following steps:

s21: encrypting vectors by using an encryption function by using the client, encrypting the original data set piece by using an encryption tool, and then transmitting the encrypted data set, the encryption tool function, the public key and auxiliary information to the server;

let the word vector data set d= { (x) generated in step S11 ₁ ，y ₁ )，(x ₂ ，y ₂ )，...，(x _m ，y _m ) Encryption with public key M using integer vector encryption scheme as D = { (c) ₁ ，y ₁ )，(c ₂ ，y ₂ )，...，(cm，y _m ) }, wherein c _i Satisfy Sc _i ＝wx _i +e _i Where w is a large integer, e _i Random errors less than w/2;

s22: uploading the encrypted ciphertext data set and the public key M to a cloud server by the client, receiving the encrypted data set by the cloud server, and then running a homomorphic encrypted logistic regression algorithm;

s3: training a privacy-protected spam filtering model;

s4: and (5) completing training of the cloud server-side junk mail filtering model, and returning model weights.

2. The training method of the spam filter based on the homomorphic encryption of integer vectors according to claim 1, wherein the step S3 specifically comprises the following steps:

s31: setting iteration times n at a cloud server side;

s32: calculating a ciphertext gradient;

s33: updating parameters by using a gradient descent method;

s34: steps S32, S33 are repeated until the number of iterations n is reached.

3. The training method of the spam filter based on the homomorphic encryption of integer vectors according to claim 1, wherein the step S4 specifically comprises the steps of:

s41: after training is completed, the cloud server side obtains an encrypted training model, and returns a ciphertext model and model weight omega to the client side;

s42: and the client receives the ciphertext model and the model weight omega, decrypts the ciphertext model by using the private key to obtain a plaintext model, namely a trained junk mail filtering model, and predicts the data by using the model.

4. A system employing the integer vector homomorphic encryption-based spam filter training method as claimed in any one of claims 1 to 3, said system comprising:

the cloud server comprises a client and a cloud server;

the client consists of a client host group and is used for providing a training data set and encrypting the training data set;

the cloud server side is used for completing an encrypted data set junk mail filtering training model;

the client encrypts the training data set by using a public key, then sends the training data set to the cloud server, the cloud server runs a privacy protection logistic regression algorithm training model, and sends the training data set to the client after training, and the client decrypts the training data set by using the private key to obtain a plaintext model.