Achieving Flexible Local Differential Privacy in Federated Learning via Influence Functions

Abstract

The use of local differential privacy in federated learning has recently grown in popularity due to rising demands for increased privacy in machine learning scenarios. While research into local differentially private federated learning is vast, the ability for a client to change their privacy parameter $\varepsilon $ ε after training, and have that change reflected in the global model’s parameters without having to repeat the entire federated training process, is currently unexplored. In this work, we propose FLDP-FL (Flexible Local Differential Privacy for Federated Learning), a simple and efficient technique for federated learning based on influence functions that enables clients to update their privacy guarantees after training without incurring extra training overhead by either the global server or the other federation participants. We show that our influence-based approach is able to accurately estimate the change in global model parameters that would occur if the client re-randomized their data under a stricter $\varepsilon $ ε and the federated learning process was repeated. Additionally, we show that our FLDP-FL approach is able to reasonably estimate the resulting change when multiple clients update their privacy parameter $\varepsilon $ ε .