Election Coding for Distributed Learning: Protecting SignSGD against Byzantine Attacks
This addresses Byzantine attacks in distributed learning for large-scale training systems, offering a novel coding-theoretic solution with incremental improvements in robustness.
The paper tackles the problem of Byzantine failures degrading learning accuracy in distributed learning with SignSGD, proposing Election Coding to guarantee robustness with minimal communication, and confirms tolerance through deep learning experiments on Amazon EC2.
Recent advances in large-scale distributed learning algorithms have enabled communication-efficient training via SignSGD. Unfortunately, a major issue continues to plague distributed learning: namely, Byzantine failures may incur serious degradation in learning accuracy. This paper proposes Election Coding, a coding-theoretic framework to guarantee Byzantine-robustness for SignSGD with Majority Vote, which uses minimum worker-master communication in both directions. The suggested framework explores new information-theoretic limits of finding the majority opinion when some workers could be malicious, and paves the road to implement robust and efficient distributed learning algorithms. Under this framework, we construct two types of explicit codes, random Bernoulli codes and deterministic algebraic codes, that can tolerate Byzantine attacks with a controlled amount of computational redundancy. For the Bernoulli codes, we provide upper bounds on the error probability in estimating the majority opinion, which give useful insights into code design for tolerating Byzantine attacks. As for deterministic codes, we construct an explicit code which perfectly tolerates Byzantines, and provide tight upper/lower bounds on the minimum required computational redundancy. Finally, the Byzantine-tolerance of the suggested coding schemes is confirmed by deep learning experiments on Amazon EC2 using Python with MPI4py package.