GATv2-Based Artificial Intelligence Framework for Intrusion Detection in UAVCAN-Enabled Unmanned Aerial Vehicle Networks

Abstract

Network-level attacks targeting unmanned aerial vehicles (UAVs) that use the UAVCAN (DroneCAN) protocol are becoming increasingly attractive due to the Controller Area Network (CAN) bus's broadcast nature. We apply and empirically evaluate Graph Attention Network (GATv2) on the task of binary intrusion detection classification over the nine scenarios of the publicly available UAVCAN attack dataset, including Flooding (scenarios 1–2), Fuzzy (scenarios 3–5), Replay (scenario 6), and compound multi-vector attacks (scenarios 7–9). Under the high-density Flooding attack in Scenario 1, the model achieves a peak weighted F1-score of 99.80% and an accuracy of 99.66%, and its performance remains above 98% across five other scenarios. We observe performance degradation under long-term Fuzzy injection (Scenario 5: accuracy 90.57%) and covert Replay attacks (Scenario 6: accuracy 91.93%), which exhibit two distinct failure modes: false-positive-dominant under statistical boundary blur and false-negative-dominant under structurally secure traffic. The performance path across attack types and compositional complexity is quantified in a comparison table for each scenario. The findings establish GATv2 as a strong baseline for graph-based UAV IDS and highlight six focused recommendations to refine the model toward full-scenario robustness.