Toward Free-Riding Attack on Cross-Silo Federated Learning Through Evolutionary Game

In cross-silo federated learning (FL), due to the heterogeneous participants, free-riders can utilize information asymmetry to make profits without performing any local model training. Free-riding attack poses possibilities and opportunities for unfairness and can seriously impair the operation of the FL ecosystem. It motivates our work to explore and characterize the unique features of free-riding attack, which differ from other attacks such as poisoning attacks. In this paper, we propose an evolutionary public goods game-based incentive model (Fed-EPG), which makes the first attempt to construct the interaction model among the participants through the evolutionary public goods game. Specifically, we consider both the public good characteristics of cross-silo FL models as well as the bounded rationality and incomplete information of competitors. We first introduce asymmetric environmental feedback to represent reward and punishment strategies in evolutionary game, and then adopt a multi-segment nonlinear control method to dynamically adjust the rewards and punishments among the participants, which achieves the incentive for the participants to cooperate stably during the training process. Experimental results validate that our incentive model is effective in the mitigation of free-riding. attacks.