A Novel Privacy Enhancement Scheme with Dynamic Quantization for Federated Learning

Federated learning (FL) has been widely regarded as a promising paradigm for privacy preservation of raw data in machine learning. Although, the data privacy in FL is locally protected to some extent, it is still a desideratum to enhance privacy and alleviate communication overhead caused by repetitively transmitting model parameters. Typically, these challenges are addressed separately, or jointly via a unified scheme that consists of noise-injected privacy mechanism and communication compression, which may lead to model corruption due to the introduced composite noise. In this work, we propose a novel model-splitting privacy-preserving FL (MSP-FL) scheme to achieve private FL with precise accuracy guarantee. Based upon MSP-FL, we further propose a model-splitting privacy-preserving FL with dynamic quantization (MSPDQ-FL) to mitigate the communication overhead, which incorporates a shrinking quantization interval to reduce the quantization error. We provide privacy and convergence analysis for both MSP-FL and MSPDQ-FL under non-i.i.d. dataset, partial clients participation and finite quantization level. Numerical results are presented to validate the superiority of the proposed schemes.