Differentially Private One-Bit Model Aggregation in Personalized Federated Learning

This paper jointly addresses communication overhead and privacy leakage in personalized federated learning (FL). We propose a model transmission method employing one-bit stochastic quantization and maximum likelihood estimation for parameter aggregation. Our approach dynamically adjusts the step size of the aggregated model updates, improving training stability of large neural networks under heterogeneous data distributions. The FL system with the proposed aggregation method is theoretically proven to achieve ($\epsilon, 0$)-differential privacy, and our convergence analysis demonstrates that performance degradation caused by one-bit transmission and privacy protection asymptotically diminishes as the FL system scales. Experiments on FMNIST and CIFAR-10 datasets validate our method’s superior accuracy and training stability compared to existing one-bit methods while preserving privacy.