Constrained Differentially Private Federated Learning for Low-Bandwidth Devices
Constrained Differentially Private Federated Learning for Low-bandwidth Devices
AI summary
A constrained federated learning scheme based on differential privacy restricts model learning to a small number of selected weights; in experiments, uplink and downlink bandwidth were reduced by up to 99.9% compared with standard federated learning. The scheme provides a theoretical privacy guarantee of differential privacy to mitigate the federated learning training-data leakage revealed by recent attacks, making it suitable for low-bandwidth devices such as mobile systems.