Abstract
Distributed approaches have many computational benefits, but they arevulnerable to attacks from a subset of devices transmitting incorrectinformation. This paper investigates Byzantine-resilient algorithms in adecentralized setting, where devices communicate directly with one another. Weleverage the so-called dual approach to design a general robust decentralizedoptimization method. We provide both global and local clipping rules in thespecial case of average consensus, with tight convergence guarantees. Theseclipping rules are practical, and yield results that finely characterize theimpact of Byzantine nodes, highlighting for instance a qualitative differencein convergence between global and local clipping thresholds. Lastly, wedemonstrate that they can serve as a basis for designing efficient attacks.