Federated Learning under Attack: Game-Theoretic Mitigation of Data Poisoning

Federated Learning (FL) is vulnerable to various attacks, including data poisoning and model poisoning, which can degrade the quality of the global model. Current solutions are based on cryptographic primitives or byzantine aggregation but suffer from performance and/or quality worsening and are even ineffective in case of data poisoning. This work proposes a game-theoretic solution to identify malicious weights and differentiate between benign and compromised updates. We use the Prisoner's Dilemma and Signaling Games to model interactions between local learners and the aggregator, allowing a precise evaluation of the legitimacy of shared weights. Upon detecting an attack, the system activates a rollback mechanism to restore the model to a safe state. The proposed approach enhances FL robustness by mitigating attack impacts while preserving the global model's generalization capabilities.