AdaDpFed: A Differentially Private Federated Learning Algorithm With Adaptive Noise on Non-IID Data

Zirun Zhao, Yi Sun, Ali Kashif Bashir, Zhaowen Lin

Abstract: The popularity of emerging consumer electronics, such as Mobile phones, PADs, and various smart home appliances, brings unprecedented convenience to people. Currently, how to obtain targeted information without revealing personal privacy is the new challenge with the application of precision recommendation and other Artificial Intelligence technologies in consumer electronics. Several solutions, based on Federated Learning transfer and update models by intermediate parameters instead of individual data, can perform some recommending tasks with weak security. However, they still face various attacks like model inversion attacks and membership inference attacks, especially in the complex non-independent and identically distribution (non-IID) environment. In this paper, we propose AdaDpFed, an adaptive federated differential private protocol in the non-IID setting. AdaDpFed can adaptively adjust the perturbation parameters, aggregated clients and sampling size according to the changing distribution of the individual data. The convergence proof shows that AdaDpFed has $\mathcal {O}\left({{}{}\frac {1}{T}}\right)$ convergence rate. Comparative experiments demonstrate that the performance of AdaDpFed outperforms other state-of-the-art protocols in both accuracy and global privacy budget.