Privacy-Preserving Federated Learning with Lightweight Transformers for UAV Intrusion Detection in IoT Networks
Subhram Dasgupta, Kushal Badal, Swetha Chittam, Xiaohong Yuan, Kaushik Roy · 2025
As UAVs become increasingly integrated into critical IoT infrastructures for applications ranging from surveillance to delivery services, they present an expanded attack surface that traditional cybersecurity frameworks are ill-equipped to handle. The distributed and resource-constrained nature of UAV-IoT networks creates a unique security paradigm where conventional intrusion detection approaches fail to provide adequate protection. The integration of Unmanned Aerial Vehicles (UAVs) into Internet of Things (IoT) ecosystems has exposed significant vulnerabilities beyond what traditional intrusion detection can manage. Current methods fail in three critical ways: centralized telemetry collection creates privacy risks, the limited processing power on UAVs cannot support heavy detection models, and rigid architectures require complete redeployment whenever threats evolve. To address these limitations, we present an intrusion detection framework that combines federated learning principles with transformer models specifically adapted for UAV-IoT constraints. Our approach employs multiple stages: we leverage SHAP analysis alongside XGBoost for interpretable feature selection, reducing computational demands by 42.3%. Additionally, we implement a VAE-GAN structure to handle the persistent issue of imbalanced attack data, while deploying a modified MobileBERT architecture that operates within strict hardware limits-requiring only 94MB storage and 136MB of active memory. Through federated learning, individual UAVs can jointly develop detection capabilities without sharing sensitive operational data, thus maintaining privacy across the network. We evaluated our framework using 54,773 UAV security samples covering five attack categories. Through VAE-GAN synthesis, we expanded the dataset to 250,000 entries to address class imbalance. The federated model reached 95.27% detection accuracy, marginally trailing the 95.42% centralized benchmark. Detection performance remained strong across attack types-DoS, false data injection, replay, and evil twin attacks all showed F1-scores exceeding 0.90. This demonstrates that UAV-IoT security may safeguard privacy without compromising performance.