Lightweight and Privacy-Enhanced Detection Model on Aerial Imagery for Post-Disaster Building Damage Reconnaissance

Abstract

As post-disaster aerial imagery becomes a crucial resource for structural damage assessment, automated detection systems must address challenges in classification granularity, data privacy, and deployment efficiency. To tackle these issues, we propose a lightweight and privacy-enhanced building damage detection framework that integrates YOLO-based object detection with differentially private training and structured pruning. Specifically, we apply Differentially Private Stochastic Gradient Descent (DP-SGD) to inject calibrated Laplace noise during training, offering formal $\varepsilon$-differential privacy guarantees for sensitive imagery. To enable real-time inference on edge-constrained platforms like UAVs, we further employ structured channel pruning to eliminate redundant parameters without modifying the model architecture. Empirical results on a well-annotated dataset demonstrate that our method maintains strong detection performance while achieving both privacy protection and model compactness, providing a lightweight and secure solution for timely post-disaster building damage assessment.