Privacy Protection in Federated Learning Based on Differential Privacy and Mutual Information
Shibo Wang, Yunbo Li, Aowei Zhao, Qiangmin Wang · 2021 3rd International Conference on Artificial Intelligence and Advanced Manufacture · 2021
Federated learning is a distributed machine learning framework. On the premise of ensuring the legitimacy and compliance of data and the security of user privacy, it realizes the joint modeling of multi-party computer groups, in order to improve the accuracy of model fitting. Aiming at the privacy security of Federated learning user data, this paper uses Laplace distributed noise and weighted aggregation algorithm with mutual information as scoring mechanism to protect user data privacy. During the experiment, the research group realized the protection algorithm proposed in the paper based on the fat platform and obtained the optimal noise parameters through multiple groups of experiments, so as to minimize the impact of the protection mechanism on the model fitting accuracy and protect the privacy and security of the data to the greatest extent. In addition, the research group reproduces the previous attack algorithm and defense Algorithm for comparative experiments, so as to verify the security and accuracy of this algorithm.