ANCHOR: A Multilayer Reputation-Based Defense for Federated Learning in Edge-Fog-Cloud IoT Systems

Multilayer federated learning (multilayer FL) improves scalability and communication efficiency in large-scale IoT networks. However, it remains vulnerable to poisoning attacks in which adversaries corrupt local training to bias the global model. Defense is particularly challenging in multilayer architectures because intermediate aggregation masks individual client behavior, and non-IID data naturally induces heterogeneous updates that can resemble malicious deviations. We propose ANCHOR, a reputationbased defense for a client-edge-fog-cloud multilayer FL architecture. ANCHOR designates a small fixed set of benign anchor STAs whose updates are used to form stable trusted references for validating the direction and magnitude of the received updates. Using these references, ANCHOR performs tier-wise verification and maintains temporal reputation scores to filter unreliable updates at the edge, fog, and cloud layers, thereby limiting the upward propagation of poisoned contributions. We evaluate ANCHOR on MNIST, Fashion-MNIST, and an air-quality classification dataset under random and biased poisoning, varying adversarial intensities, and both IID and non-IID data distributions. Results show that ANCHOR preserves benign-learning performance comparable to a no-attack baseline, while achieving superior robustness and detection effectiveness relative to existing defenses under adversarial settings, with lower defense-andaggregation execution time in the multilayer pipeline.