Spectre-Fed: Evolving federated edge intelligence from FedEdge-ID to robust-private IoT intrusion detection via hybrid adversarial training

Ullah, S., J, W. U., Kamal, M. M., Alzaylaee, M. K., & Alibakhshikenari, M. (2026). Spectre-Fed: Evolving Federated Edge Intelligence From FedEdge-ID to Robust-Private IoT Intrusion Detection via Hybrid Adversarial Training. IEEE Open Journal of the Communications Society, 7, 1994-2012. https://doi.org/10.1109/OJCOMS.2026.3665325

Abstract

The growing number of Internet of Things (IoT) devices requires decentralized Edge Intelligence solutions. As the current FL-based IDS systems are decentralized solutions for privacy protection, face two major problems: 1) network traffic manipulation through adversarial evasion attacks 2) privacy threats from gradient-based inference attacks and 3) server-side Robustness issue. The current methods which use Differential Privacy (DP) or adversarial training result in 5-15% accuracy reduction which makes them unsuitable for deployment. The key novelty of our work is the integration of a novel dual-defense framework that uniquely reconciles the conflict between differential privacy noise and adversarial gradient requirements, effectively eliminating the conventional “accuracy tax” along with server-side Robust Aggregation. Our research develops an enhanced two-stage federated system which is robust and protects privacy while delivering secure IoT edge intelligence solutions. The core system FedEdge-ID provides 99.73% detection performance across different edge devices. Spectre-Fed enhances the FedEdge-ID framework via three key defenses: (1) Hybrid Loss Adversarial Training ( α =0.5) to fortify decision boundaries against evasion, (2) Gradient-Guided Adaptive Privacy with decreasing noise injection ( σ0 =0.0005, γ =0.95) for secure gradient updates, and (3) Robust Trimmed Mean Aggregation to counter Byzantine poisoning. Experiments demonstrate that Spectre-Fed’s client-side (Layer 1) defense achieves 99.72% clean accuracy with only a 0.01% utility loss versus the non-private baseline. It shows strong adversarial resilience, retaining 99.34% accuracy against FGSM attacks ( ϵ =0.01), a mere 0.38% degradation from the clean state. When integrated with server-side Robust Aggregation (Layer 2), the system sustains 99.59% accuracy even under active label-flipping attacks from 20% of clients, while preserving high utility compared to the baseline. The system achieves optimal privacy-utility balance through its formal privacy protection and its ability to resist adversarial attacks which makes it suitable for zero-trust IoT systems.

Funder

European Union
European Research Executive Agency
Taighde Éireann - Research Ireland
University of Galway

Publisher

Institute of Electrical and Electronics Engineers

Rights

CC BY

cb

Collections

Publications