Changliang Zou, Professor, Nankai University

Xin Liu, Professor

Decentralized Byzantine Machine Identification

8:00-9:00(Friday) September 19, N714

We present a p-value free, dimensionally insensitive method for decentralized Byzantine machine detection and stochastic optimization. Our detection scheme uses sample-splitting and a data-driven threshold to achieve a finite-sample bound on the false discovery rate and a sure-detection guarantee. In the optimization phase, we reconstruct a row-stochastic mixing matrix and prove that our stochastic DGD-RS algorithm attains an O(1/√K) non-asymptotic convergence rate for nonconvex functions, ensuring exact convergence with high probability. The effectiveness of our approach is demonstrated through extensive simulations.