Federated Learning Gradient Encryption for Financial Data Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing federated learning methods for financial data sharing face challenges in ensuring data privacy, as gradient parameters can leak during the training process, violating data security regulations.
Innovation Solution
A privacy protection method and system that employs multi-key homomorphic encryption to encrypt and decrypt gradient parameters, ensuring they remain local, using a cloud server, master servers, edge servers, and clients to securely aggregate and update model parameters without data exchange between institutions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If gradient parameters are uploaded during federated learning training, then model training can be performed across multiple institutions, but data leakage occurs compromising privacy protection
Solution Approach 1:
The patent introduces homomorphic encryption as an intermediary mechanism that allows gradient parameters to be processed (aggregated) in encrypted form. The encryption scheme acts as a mediator between the need for parameter aggregation and the requirement for privacy protection, enabling the aggregation of encrypted gradients without decrypting them first, thus preventing data leakage while maintaining training functionality
Solution Approach 2:
The patent transforms the gradient parameters from their original plaintext form to encrypted form through homomorphic encryption. This parameter change allows the system to operate on encrypted data, fundamentally altering the state of the parameters to provide privacy protection while still enabling mathematical operations necessary for federated learning aggregation
2Productivity
If data is integrated from different organizations for joint training, then modeling capability is improved, but data security and privacy regulations are violated
Solution Approach 1:
The patent segments the centralized data integration process into distributed federated learning across multiple institutions. Each institution keeps its data locally segmented and only shares encrypted gradient updates, eliminating the need to centralize sensitive data while still achieving joint modeling capabilities through coordinated encryption-based aggregation
Solution Approach 2:
Homomorphic encryption serves as an intermediary that enables data integration from different organizations without direct data sharing. The encryption mechanism allows the system to combine information from multiple sources while maintaining regulatory compliance by ensuring that raw data never leaves its originating organization
3Loss of energy
If only gradient parameters are uploaded for federated learning, then communication overhead is reduced, but privacy protection is insufficient due to potential leakage
Solution Approach 1:
The patent changes the state of gradient parameters from plaintext to encrypted form before transmission. This parameter transformation maintains the compact size of gradient updates (preserving low communication overhead) while adding cryptographic protection that prevents privacy leakage during transmission and aggregation
Data Source
AI summary
A privacy protection method and system for financial data sharing based on federated learning are provided. In recent years, due to the restrictions of data security and privacy protection laws and regulations, it is difficult to share data across institutions or departments. In order to make data transfer and transaction between different entities can be achieved without violating the national laws on data privacy and data security, the privacy protection method and system for financial data sharing based on federated learning is provided. A privacy collection intersection technology is adopted, so that two institutions, which may have many differences in business, but most of their customer groups are the same, jointly train a learning model.