AeroVeil-FL: Privacy-preserving federated learning framework for UAV-assisted real-time disaster detection

Unmanned Aerial Vehicles (UAVs) have become essential tools for disaster monitoring and response due to their quick deployment and high-resolution sensing capabilities. However, collecting and sharing UAV data pose significant risks to location privacy, which can threaten both operational security and sensitive information. This study introduce a novel federated learning (FL) framework which combined with multi-hop mix networks and adaptive differential privacy, to safe UAV location data during disaster detection. The framework enables UAVs to collaboratively build a global model without sharing raw data, thereby protecting sensitive positional information from potential adversaries. Simulation shows that the proposed method can reach up to 95% accuracy in the global model, even with highly diverse UAV data, surpassing benchmark results by 5%–12%. The loss function converges from 0.82 to 0.12 over 50 epochs, achieving faster convergence by 20%–40% compared to traditional methods. The adversary’s error in estimating position increases to 0.9 under the proposed privacy computation, which shows a 5%–10% improvement. Additionally, model accuracy remains stable across various noise distributions, maintaining 2%–6% higher accuracy. These results show that the proposed technique effectively balances location privacy, model performance and convergence efficiency.