Mitigating Backdoor Attacks in Federated IoT Systems via Representative-Attention Mechanism

The heterogeneous nature of IoT environments produces diverse, non-independent, and identically distributed (non-IID) data, making the detection of backdoor attacks in federated learning settings more challenging. Traditional defenses, relying on model weights or gradient-based statistical deviations, make restrictive assumptions about attack methods and client data distributions, hence limiting their effectiveness. In this paper, we propose FeRA (Federated Representative-Attention), a novel defense mechanism that leverages cross-client attention over internal feature representations to distinguish benign from malicious clients. Our evaluation demonstrates FeRA’s robustness across various federated learning scenarios, including challenging non-IID data distributions typical of IoT deployments.