Reputation-driven Cooperation in Lattice-based Decentralized Federated Learning through Evolutionary Game Theory

The paper proposes an evolutionary game-theoretic framework for decentralized federated learning (DFL). It models peer-to-peer interactions on a lattice under bounded rationality, incorporates training costs, communication overhead, cooperative rewards, and a reputation-based reward-and-punishment mechanism. According to the abstract, simulations raise average accuracy from roughly 70% to 82%, increase cooperation frequency from below 5% to nearly 100%, and reduce accuracy variance from about 0.40 to 0.002 relative to a baseline. The abstract does not provide the underlying datasets, model architectures, parameter settings, or full comparison protocol.