Decentralized ML Resilience Under Adversarial Majority

A new study from arXiv (2605.07841) proposes an incentive-based framework for decentralized machine learning when adversaries control a majority of worker nodes. Existing robust aggregation methods assume honest-majority settings and fail under adversary-dominated conditions. The framework rewards reports only when they are mutually consistent up to a threshold, transforming adversaries into rational agents balancing estimation error against reward loss. The work examines iterative optimization, highlighting a trade-off between permissive acceptance rules that enable faster early progress but admit more corruption, and strict rules that improve accuracy but cause frequent rejections.