distributed-learning

This paper introduces 'dictator clients'—a novel class of malicious participants in federated learning capable of erasing other clients' contributions while preserving their own—and provides theoretical analysis of their impact on model convergence, including scenarios with multiple adversarial clients.