SLEIGHT: Real-time Closed-Loop Attack Analytics Framework for Holistic Intrusion Detection Systems in Unmanned Aerial Vehicles

Unmanned aerial vehicles (UAVs) are increasingly secured by advanced intrusion detection systems (IDSs) that monitor the holistic sensor–control–actuator pipeline. However, the robustness of these IDSs has not been systematically characterized against adaptive, real-time, physically plausible, and stealthy telemetry perturbations during closed-loop flight. Prior adversarial studies often rely on open-loop, offline testing, failing to capture the temporal, dynamic, and closed-loop constraints inherent to UAV operation. Consequently, it remains unclear what mission-level impact an in-flight data-stream attacker can induce while maintaining benign-looking behavior. We propose SLEIGHT, a novel, real-time, attack analytics framework that maps the stealth–impact frontier of closed-loop UAV IDSs under physically plausible perturbations. SLEIGHT trains a phase-conditioned reinforcement-learning policy with attacker-side plausibility enforcement to synthesize bounded perturbations maximizing mission impact under stealth and feasibility constraints. Using Pixhawk hardware- and software-in-the-loop experiments with MAVSDK, SLEIGHT benchmarks four IDSs, including state-of-the-art controller- and actuator-aware holistic IDSs such as PID-Piper and ConFIDe. SLEIGHT consistently achieves high no-alarm rates across diverse and multi-UAV missions (>96% in single-channel GPS, sustaining >95% and >90% against ConFIDe and PID-Piper, respectively) while inducing trajectory drift and landing displacement up to ≈11 m in single UAV HITL and > 60 m in swarm scenarios.