IRS-Aided Federated Learning with Dynamic Differential Privacy for UAVs in Emergency Response

Abstract

The unforeseen events of natural disasters often devastate critical infrastructure and disrupt communication. The use of unmanned aerial vehicles (UAVs) in emergency response scenarios offers significant potential for delivering real-time information and assisting emergency response efforts. However, challenges such as physical barriers to communication not only hinder transmission performance by obstructing established line-of-sight (LoS) links but also pose risks to the privacy of sensitive information exchanged across these links. To address these challenges, we propose a novel IRS-aided UAV secure communications framework aimed to enhance communication efficiency while ensuring privacy preservation in emergency response scenarios. The framework consists of three stages: (i) local model training with dynamic differential privacy mechanism using stochastic gradient descent (SGD), with adaptive learning rate adjustment based on validation performance, (ii) decentralized federated learning (FL) with intelligent reflective surfaces (IRS) incorporation to improve communication and information exchange between UAV-to-UAV and UAV-to-ground station, and (iii) selection of a UAV header based on operational characteristics and connectivity to aid UAV-to-ground station communication.Furthermore, we evaluated our proposed framework through experimental simulations and achieved 0.91 accuracy after 50 federated learning rounds underscoring the efficacy of our dynamic noise and learning rate adjustment mechanism. Additionally, our integration of IRS led to lower communication latency, highlighting the effectiveness of our approach. This framework adeptly balances privacy protection with model accuracy.