SHIELD-FL: Scalable backdoor defense in federated learning via gradient trust and data-free distillation under non-IID data

Federated Learning (FL) enables collaborative model training without sharing raw data but remains vulnerable to backdoor attacks, particularly under high adversarial ratios and non-IID data distributions. Existing defenses often rely on clean public datasets or strong threat assumptions, limiting real-world applicability. We propose SHIELD-FL (Secure Hybrid Inspection for Enhanced Learning Defense in Federated Learning), a scalable, attack-agnostic framework that operates without external data. SHIELD-FL integrates: (1) Gradient Trust Indexing, which dynamically scores client reliability via adversarial perturbation sensitivity; (2) Adaptive Clustering using HDBSCAN in parameter space to isolate benign clients; and (3) Robust Knowledge Distillation with temperature-scaled soft labels and stochastic weight averaging. Extensive evaluations on CIFAR-10, EMNIST, and Fashion-MNIST under five adaptive backdoor attacks show SHIELD-FL achieves up to 92.5% main-task accuracy while reducing attack success rates to 3.6%, even with 60% malicious clients. It outperforms data-dependent defenses like FLTrust, maintains low communication overhead, and runs 3–4× faster than ensemble methods. SHIELD-FL is especially suitable for privacy-sensitive, resource-constrained environments including emerging MENA region applications.