Privacy-Preserving Federated Learning with Adaptive Differential Privacy in Distributed Networks

Abstract

Federated learning lets many devices or institutions train a shared model without moving their raw data to a central location, yet the model updates they send can still reveal sensitive information about the data on which they were computed. Differential privacy offers a rigorous defense by clipping each update and adding calibrated random noise, but most deployments use a single fixed noise level and a single fixed clipping bound for every client, every round and every layer of the model. This is wasteful. It adds more noise than needed when updates are small and late in training, it ignores the fact that clients differ in how sensitive their data are, and it makes accuracy fall sharply when strong privacy is required. This paper proposes an adaptive differential privacy framework for federated learning in distributed networks.