Lightweight One-Class Autoencoder for UAV Intrusion Detection

Abstract

Unmanned aerial vehicles (UAVs) operate in dynamic and resource-constrained environments, requiring lightweight and reliable intrusion detection methods. However, existing approaches often rely on labeled attack data or computationally intensive models, limiting their practicality under onboard computational constraints. To address this issue, we propose a lightweight UAV intrusion detection framework based on a one-class autoencoder with a denoising and simplified self-supervised representation learning (SSRL) mechanism. The framework is trained using only benign data and integrates sensor-wise preprocessing, Principal Component Analysis (PCA), and a compact autoencoder architecture for efficient anomaly detection. A quantile-based threshold is applied to reconstruction error to enable stable and low-cost decision-making, avoiding complex adaptive strategies. Experimental results demonstrate that the proposed method achieves competitive detection performance while improving robustness to noise and significantly reducing inference time and memory overhead, indicating strong potential for deployment-oriented UAV intrusion detection under resource-constrained environments.