Byzantine-Robust and Differentially Private Federated Optimization Under Weaker Assumptions
arXiv·medium signal
Islamov et al. achieve Byzantine fault tolerance and differential privacy simultaneously in federated learning without the strong assumptions (bounded gradients, IID data) required by prior work. The method handles heterogeneous clients with adversarial participants while preserving privacy guarantees. Practical for any federated/distributed ML system where both security and privacy are requirements.