Encrypted Gradient Exchange for Joint Training Data Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In joint training of machine learning models, there is a challenge in ensuring data security while maintaining model performance by preventing information leakage between participants.
Innovation Solution
A data protection method involving encryption of identification information using keys from both active and passive participants, and determining generation gradient information based on actual gradient information, generation label information, or feature information of passive participants to update the joint training model parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If joint training model is used to complete machine learning training without exposing original data, then data security is improved, but model performance may deteriorate due to information isolation between participants
Solution Approach 1:
The patent introduces encrypted gradient information as an intermediary that enables collaboration between participants without direct data exposure. The active participant generates gradient information that is encrypted using both its own key and the passive participant's key, allowing the passive participant to verify and use the gradient information without the active participant seeing the original data, thus maintaining both security and model performance
Solution Approach 2:
The patent transforms the training process by changing the parameter representation from raw data to encrypted gradient information. By encrypting the gradient information with dual keys and transforming it into a verifiable format, the system allows model updates while maintaining data security, resolving the contradiction between security and performance
2Reliability
If encryption is applied to prevent information leakage between participants, then data privacy is improved, but system complexity increases due to key management and encrypted computation
Solution Approach 1:
The patent segments the encryption process into distinct phases: the active participant encrypts gradient information with its own key, then the passive participant adds its key layer. This segmentation allows each participant to manage only their own key without needing to handle others' keys directly, reducing overall system complexity while maintaining strong privacy protection
Solution Approach 2:
Each participant independently manages their own encryption key and performs encryption/decryption operations on their own data. The active participant self-encrypts gradient information, and the passive participant self-decrypts using their own key, eliminating the need for a centralized key management system and reducing system complexity
3Reliability
If generated data is used to protect privacy in joint training, then information leakage is prevented, but model training accuracy may deteriorate due to synthetic data limitations
Solution Approach 1:
The patent creates encrypted copies of gradient information that preserve the statistical properties needed for model training while removing direct data exposure. The encrypted gradient information serves as a faithful copy that maintains training accuracy because it is derived from real data through mathematical transformations rather than synthetic generation, thus preventing information leakage without sacrificing precision
Data Source
AI summary
The present disclosure relates to a data protection method, apparatus, medium and electronic device. The method comprises: obtaining a specified batch of reference samples of an active participant of a joint training model; determining generation gradient information of the first reference sample; determining target gradient information sent to the passive participant according to the generation gradient information, and sending the target gradient information to the passive participant, to update, by the passive participant, parameters of the joint training model according to the target gradient information. Through the above solution, the influence of the generated data on the training process and model performance of the joint training model is avoided as much as possible, and the privacy and security of data are improved.