Subject
3 entries
Distributed Ml
Bookmarks
Secure Byzantine-Robust Machine Learning
He, Karimireddy, and Jaggi propose a two-server cryptographic protocol that simultaneously achieves Byzantine robustness, input privacy, and local differential privacy for distributed machine learning — three properties usually addressed separately. Bridges the privacy-robustness gap in federated learning.
Learning from History for Byzantine Robust Optimization
Karimireddy, He, and Jaggi (EPFL, arXiv:2012.10333, 2021) propose using historical gradient information to detect and filter Byzantine workers in distributed training, achieving near-optimal convergence even with a constant fraction of corrupt workers. The historical approach is notable because it breaks the fundamental limitation of single-round Byzantine filters without requiring cryptographic overhead.
Asynchronous Decentralized SGD with Quantized and Local Updates
SwarmSGD is an asynchronous decentralized optimization algorithm that provably converges when combining gossip communication, gradient quantization, and local update steps simultaneously across heterogeneous data distributions. It achieves performance comparable to large-batch SGD on supercomputing systems while reducing communication overhead.