Compression Improves Differentially Private Federated Learning

Compressed sensing can improve the model quality of differentially private federated learning without sacrificing privacy. Differential privacy defends against inference and reconstruction attacks based on intercepted gradients by adding noise to the exchanged update vectors, but the amount of noise is proportional to model size, and when the model is too large this leads to quality degradation. Experiments on 2 datasets show that the scheme reduces communication cost by up to 95%, with negligible performance loss compared with conventional non-privacy-preserving federated learning.