Research Article
A Scheme of Anti Gradient Leakage of Federated Learning Based on Blockchain
@INPROCEEDINGS{10.1007/978-3-031-30623-5_7, author={Xin Zhang and Yuanzhen Liu and Yanbo Yang and Jiawei Zhang and Teng Li and Baoshan Li}, title={A Scheme of Anti Gradient Leakage of Federated Learning Based on Blockchain}, proceedings={Security and Privacy in New Computing Environments. 5th EAI International Conference, SPNCE 2022, Xi’an, China, December 30-31, 2022, Proceedings}, proceedings_a={SPNCE}, year={2023}, month={4}, keywords={Federated Learning Blockchain Ring signature Deep Leakage from Gradients Gradient Update}, doi={10.1007/978-3-031-30623-5_7} }- Xin Zhang
Yuanzhen Liu
Yanbo Yang
Jiawei Zhang
Teng Li
Baoshan Li
Year: 2023
A Scheme of Anti Gradient Leakage of Federated Learning Based on Blockchain
SPNCE
Springer
DOI: 10.1007/978-3-031-30623-5_7
Abstract
Federated learning provides a new solution for data security and privacy protection in the process of machine learning. In distributed learning, interactive gradient rather than direct exchange of data avoids the direct acquisition of data by malicious participants. The latest research shows that the gradient will also leak the data privacy during user training, current solutions to gradient leakage problems are mainly divided into two categories: 1. Encrypt the parameters of the model updating process by using the cryptography scheme; 2. Increase the interference noise for the gradient through thinning, gradient compression, adding noise and other schemes. However, the cryptographic scheme cannot be targeted at the aggregator, and the noise scheme has too much impact on the performance of the model. This paper uses the trust mechanism of blockchain to design an anti gradient leakage scheme for this problem: when the participants upload the gradient, the noise is added through the blockchain smart contract, and the noise value is randomly divided into multiple copies and stored in the blockchain. The central server obtains the noise sum of all participants through the smart contract, and aggregates the gradient to remove the noise. In the training process, all participants can only obtain the aggregation gradient and noise gradient, and cannot recover user data. At the same time, due to the automatic execution and tamper proof characteristics of blockchain smart contracts, malicious interference in the data exchange process is avoided.
