Resource-Constrained Secure Federated Learning Framework for Unmanned Aerial Vehicles

Abstract

Unmanned Aerial Vehicles (UAVs) are deployed in mission-critical applications such as surveillance, disaster relief, and ecological monitoring. These use cases demand real-time intelligence from sensitive data, necessitating privacy-preserving distributed learning approaches like Federated Learning (FL). However, conventional secure FL mechanisms are impractical for UAV environments due to their limited battery life, computational power, and bandwidth constraints, especially those relying on computationally intensive encryption and model aggregation. The proposed Responsible Federated Learning for Unmanned Aerial Vehicles (ResFL-UAV) is a resource-constrained secure FL framework optimized for UAVs using the Stochastic Controlled Averaging for Federated Learning (SCAFFOLD) algorithm to reduce client drift due to non-IID data distributions. ResFL-UAV integrates SCAFFOLD's variance reduction strategy with gradient sparsification, Lightweight Differential Privacy (DP), and a compressed secure aggregation protocol to reduce overhead while protecting data privacy and considerably improving model performance. The proposed ResFL-UAV framework achieves 94.2% accuracy on the Airborne UAV Detection dataset and 92.8% on the Drone Detection Dataset. These findings show that ResFL-UAV, supported by SCAFFOLD, provides a scalable and highly secure FL framework for resource-limited UAV networks.