Toward Adaptive Intrusion Detection Systems for UAVs Using Cyber-Physical Image Analysis

The rapid proliferation of uncrewed aerial vehicles (UAVs) across modern applications has significantly heightened their susceptibility to cyber and physical security threats. Despite the growing body of research on intrusion detection systems (IDS) for UAVs, many existing solutions exhibit significant limitations, including a reliance on synthetic datasets and binary classification frameworks. These constraints hinder their realism and effectiveness in detecting diverse and evolving threats. Additionally, conventional approaches often separate cyber and physical features, overlooking critical interdependencies that can reveal sophisticated attacks such as GPS spoofing or packet injection. Imbalanced datasets and narrowly defined attack scopes further compromise the adaptability and robustness of these systems. To address these challenges, this study proposes a novel image-based deep learning approach that jointly analyzes cyber and physical features extracted from real UAV datasets. The framework transforms tabular sensor data into multichannel image representations using a novel Gramian Angular Field (GAF) method that we developed and named Dual-GAF, enabling high-fidelity feature extraction and contextual data consideration via Convolutional Neural Networks (CNNs). To mitigate data imbalance and enhance classification performance across multiple attack types, the Synthetic Minority Over-Sampling Technique (SMOTE) is applied during training. Comprehensive experimental evaluations were conducted across varying attack complexities and feature configurations —cyber-only, physical-only, and combined cyber-physical— using real UAV datasets. The results consistently demonstrate the superior detection performance of the proposed IDS framework. efficiency.