CN-121984654-A - Semi-homomorphic proxy re-encryption safety federal recommendation method and device

CN121984654ACN 121984654 ACN121984654 ACN 121984654ACN-121984654-A

Abstract

The invention provides a semi-homomorphic proxy re-encryption safe federation recommendation method and device, the method comprises the steps of obtaining user interaction data, inputting the user interaction data into a recommendation model to obtain recommendation results output by the recommendation model, wherein the recommendation model is obtained based on global optimal parameter configuration sent by a proxy server, global optimal parameters are obtained based on federation learning training of multiple iterations, and in each iteration, the proxy server aggregates local update project factors from a plurality of clients and safely distributes the local update project factors to each client for local training by utilizing a semi-homomorphic re-encryption technology, and receives the returned local update project factors. The method can effectively prevent the third party from stealing the model parameters after being wrongly or broken, and ensure that extremely accurate and personalized recommendation results are generated.

Inventors

SUN YI
HAN ZHIYUAN
ZHOU ZAN
XU GUANGYU

Assignees

北京邮电大学深圳研究院
北京邮电大学

Dates

Publication Date: 20260505
Application Date: 20251203

Claims (10)

1. A semi-homomorphic proxy re-encryption safety federal recommendation method is characterized in that, Acquiring user interaction data; The recommendation model is configured based on global optimal parameters sent by a proxy server, the global optimal parameters are obtained based on federal learning training of multiple iterations, in each iteration, the proxy server aggregates local update project factors from a plurality of clients, and safely distributes the local update project factors to each client for local training by utilizing a semi-homomorphic re-encryption technology, and receives the returned local update project factors.
2. The semi-homomorphic proxy re-encryption secure federal recommendation method according to claim 1, comprising, prior to entering the user interaction data into a recommendation model: Receiving a re-encryption message sent by the proxy server, wherein the re-encryption message is encrypted by the proxy server on a global project latent factor generated by initialization by using a first public key generated previously and is safely distributed by using a semi-homomorphic re-encryption technology; decrypting the re-encrypted message, and carrying out local training by combining with historical user interaction data to determine local project factors, local user factors and local scoring data, wherein the local scoring data is used for representing the interaction strength of project features in the local project update factors and user features in the local user update factors; Encrypting the local project factor by using a first public key sent by the proxy server which is received before, and sending the encrypted local project factor to the proxy server; And receiving and decrypting the re-encrypted message updated by the proxy server based on each received local project factor, updating the local project factors and the local user factors by combining the local scoring data and an alternate least squares algorithm, encrypting the updated local project updating factors by using the first public key again and sending the encrypted local project updating factors to the proxy server so as to cooperate with the proxy server to perform iterative training until the preset iteration times are reached, and receiving global optimal parameters sent by the proxy server.
3. The semi-homomorphic proxy re-encryption secure federal recommendation method according to claim 2, further comprising, prior to entering the user interaction data into a recommendation model: generating a second key pair, wherein the second key pair comprises a second private key and a second public key; sending a registration request to the key management server, and sending the second key pair to the key management server, so that the key management server generates a re-encryption key sent to the proxy server based on the second key pair and a first key pair generated in advance, wherein the first key pair comprises a first public key and a first private key; And receiving a first public key returned by the key manager based on the registration request.
4. The semi-homomorphic proxy re-encryption secure federal recommendation method of claim 3, wherein the re-encryption message comprises global item semi-homomorphic re-encryption factors and homomorphic outer-volumes for each scoring item; The global project semi-homomorphic re-encryption factor is obtained by the proxy server performing semi-homomorphic re-encryption by using a re-encryption key based on a global project encryption factor of a corresponding scoring project, the scoring project corresponding to the global project encryption factor is obtained by the proxy server aggregating local update project factors from clients based on the corresponding scoring project, and the re-encryption key is generated by the key management server based on a first private key of the first key pair and a second private key of the second key pair and is sent to the proxy server; The homomorphic external sum is obtained by the proxy server through homomorphic operation based on the global item re-encryption half homomorphic factors corresponding to all scoring items.
5. The semi-homomorphic proxy re-encryption secure federal recommendation method of claim 4, wherein the global project semi-homomorphic re-encryption factor is obtained by the proxy server based on a user set of corresponding scoring projects and combining the semi-homomorphic encryption factors of the corresponding scoring projects by using a one-level homomorphic scalar algorithm; The semi-homomorphic encryption factor is obtained by the proxy server selecting global project re-encryption factors corresponding to any two users from the user set corresponding to the scoring project and utilizing one-level homomorphic addition operation; The global project re-encryption factor is obtained by re-encrypting the corresponding global project encryption factor by the proxy server through a re-encryption algorithm.
6. The semi-homomorphic proxy re-encryption secure federal recommendation method of claim 3, wherein the re-encryption message comprises global item semi-homomorphic encryption factors and homomorphic re-encryption outer volumes for each scoring item; the global project semi-homomorphic encryption factor is obtained by the proxy server through semi-homomorphic operation based on the global project encryption factor of the corresponding scoring project; the homomorphic re-encryption outer sum is obtained by the proxy server based on each secondary re-encryption result and by utilizing a secondary homomorphic addition operation; the secondary re-encryption result is obtained by the proxy server performing secondary re-encryption by using a re-encryption key based on each homomorphic operation result, wherein the re-encryption key is generated by the key management server based on a first private key in the first key pair and a second private key in the second key pair and is sent to the proxy server; The homomorphic operation result is obtained by the proxy server through one-level homomorphic multiplication operation based on the global item semi-homomorphic encryption factors corresponding to any two scoring items in all scoring items.
7. A semi-homomorphic proxy re-encryption secure federal recommendation device, comprising: the data acquisition module acquires user interaction data; The recommendation module inputs the user interaction data into a recommendation model to obtain a recommendation result output by the recommendation model, wherein the recommendation model is obtained based on global optimal parameter configuration sent by a proxy server, the global optimal parameter is obtained based on federal learning training of multiple iterations, and in each iteration, the proxy server aggregates local update project factors from a plurality of clients and safely distributes the local update project factors to each client for local training by utilizing a semi-homomorphic re-encryption technology and receives the returned local update project factors.
8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the processor implements the semi-homomorphic proxy re-encryption secure federal recommendation method of any one of claims 1 to 6 when the computer program is executed by the processor.
9. A non-transitory computer readable storage medium having stored thereon a computer program, which when executed by a processor implements the semi-homomorphic proxy re-encryption secure federal recommendation method according to any one of claims 1 to 6.
10. A computer program product comprising a computer program which when executed by a processor implements the semi-homomorphic proxy re-encryption security federal recommendation method according to any one of claims 1 to 6.

Description

Semi-homomorphic proxy re-encryption safety federal recommendation method and device Technical Field The invention relates to the technical field of artificial intelligence, in particular to a semi-homomorphic proxy re-encryption safety federal recommendation method and device. Background In the information age, large amounts of data present information overload problems that are alleviated by the introduction of recommendation systems. The recommendation system is applied to various business scenes, and is everywhere visible in daily Internet life of people. With the increase of people's privacy protection consciousness, privacy security in user data is increasingly emphasized, and privacy protection in recommendation systems is becoming important. In the information age, the personalized recommendation system has become a core technology for improving user experience and commercial value, and the traditional implementation mode of the personalized recommendation system depends on centralized machine learning, namely, massive user behavior data are gathered to a central server for model training. However, with the general arousal of the user's awareness of personal data security, this mode of data set faces an unprecedented challenge, which not only carries a significant risk of privacy disclosure, but also limits the potential for the model to be optimized with global data due to the data islanding problem. Therefore, exploring a technical paradigm that can still effectively perform collaborative learning on the premise of protecting user privacy has become a key problem to be solved in the field of artificial intelligence. To address the above challenges, federal learning has emerged as an innovative, distributed machine learning paradigm that distributes a global initial model to each participant, primarily through a central server, each participant trains the model independently using its locally stored data, and sends updated parameters of the model back to the central server, which then fuses the model updates from multiple participants through a secure aggregation protocol, thereby generating a new global model with better performance. However, since each round of training in federal learning requires the transmission of model update gradients between a server and a large number of clients, these gradient information accurately reflect local data characteristics, indicating how model parameters should be adjusted to better fit a particular data set, once an attacker obtains multiple iterations of gradient updates, member inference attacks can be launched to determine whether a particular data sample exists in a certain client local training set, even gradient leakage attacks or model reverse attacks can be launched to reconstruct the original training data of a client with high accuracy, this ability to reverse sensitive information from model updates constitutes a fundamental potential threat to the security of the bang learning. Disclosure of Invention The invention provides a semi-homomorphic proxy re-encryption safe federal recommendation method and device, which are used for solving the privacy safety defect that the original data of a user is inferred or rebuilt due to model gradient update leakage in the prior art, effectively preventing a third party from averting or stealing model parameters after being breached, and ensuring that extremely accurate and personalized recommendation results are generated. The invention provides a semi-homomorphic proxy re-encryption safe federation recommendation method which comprises the steps of obtaining user interaction data, inputting the user interaction data into a recommendation model to obtain a recommendation result output by the recommendation model, wherein the recommendation model is obtained based on global optimal parameter configuration sent by a proxy server, the global optimal parameter is obtained based on federation learning training of multiple iterations, and in each iteration, the proxy server aggregates local update project factors from a plurality of clients and safely distributes the local update project factors to each client to carry out local training by utilizing a semi-homomorphic re-encryption technology, and receives the returned local update project factors. The semi-homomorphic proxy re-encryption safe federal recommendation method comprises the steps of receiving re-encryption information sent by a proxy server, wherein the re-encryption information is obtained by encrypting global project potential factors generated through initialization by using a first public key generated in advance and safely distributing the global project potential factors through a semi-homomorphic re-encryption technology, decrypting the re-encryption information, carrying out local training by combining historical user interaction data, determining local project factors, the local user factors and local scoring data, wherein the local scoring da