SAFER-FL: Adversarially Robust Federated Learning in IoT Networks via Latent Space Auditing and Verifiable Contributions

The rise of IoT networks has driven demand for robust, privacy-preserving machine learning in distributed, resource-constrained environments. Federated Learning (FL) addresses this by enabling on-device training without sharing raw data, but remains vulnerable to data poisoning from adversarial clients. Ensuring auditable data quality and verifiable model contributions is essential for securing FL and challenging due to limited client data visibility. Existing defenses often rely on trusted central servers for client selection or aggregation, offer limited visibility into client data quality and often fail to identify clients contributing low-quality or poisoned data. To overcome these limitations, we propose Statistical Auditing and Filtering for Enabling Robust FL (SAFER-FL), a data-aware framework for adversarially robust FL in IoT networks. SAFER-FL builds a global, privacy-preserved statistical profile by aggregating Gaussian-noised client-side data moments (mean, covariance) in latent space. Clients use this to filter poisoned samples via Mahalanobis distance before training, ensuring only clean data influences the model. This proactive approach eliminates the need for validation data or pretrained models. We provide theoretical analysis on convergence, filtering efficacy, and poisoning resilience. Experiments on five IoT datasets show SAFER-FL significantly improves robustness and accuracy under various attacks, making it practical for real-world IoT FL deployments.