CODP: Improving Differentially Private Federated Learning by Cascading and Offsetting Noises Between Iterations

Federated learning (FL) has attracted tremendous attention due to its capability to preserve data privacy. In FL, a parameter server (PS) without accessing clients' raw data can assist decentralized clients in completing model training by aggregating and distributing model parameters for multiple iterations. From the clients' perspective, exposing model parameters to the PS can still result in privacy leakage. To further enhance privacy protection, differentially private federated learning (DPFL) is invented, in which clients add differentially private (DP) noises to distort their parameters to be exposed. However, the main challenge of DPFL lies in inferior model accuracy due to DP noises. To overcome this challenge, in this paper we propose a novel algorithmic framework for DPFL, which is called CODP, by cascading and offsetting DP noises between iterations. In existing works, each DPFL client only considers how to protect its model parameters based on the number of iterations to expose parameters overlooking the underlying relation of model parameters in consecutive iterations. The novelty of CODP lies in cascading DP noises from each iteration to its subsequent iteration so that DP noises can be offset in the subsequent iteration, and hence model accuracy can be improved. Additionally, we theoretically prove that CODP can substantially improve the convergence rate of DPFL without compromising privacy preservation by leveraging the most widely used Laplace and Gaussian mechanisms, respectively. We conduct comprehensive experiments using MNIST, Fashion-MNIST, and Lending Club datasets to demonstrate that the model accuracy of DPFL can be remarkably improved by CODP with a fixed privacy budget.