Ensemble Learning-Based Intrusion Detection System for Aerial Base Stations Against Adversarial Evasion Attacks

Aerial base stations (ABSs) are expected to play a major role in 5G+ wireless networks where unmanned aerial vehicles (UAVs) serve as flying base stations to provide wireless coverage. As the utilization of ABSs continues to expand, ensuring their security and resilience against malicious attacks emerges as a vital concern. In this paper, we demonstrate successful false data injection attacks on a real UAV testbed, leading the UAV offcourse, which can impact the ABS coverage. Current research efforts focus on designing intrusion detection systems (IDS) tailored specifically for UAVs. However, the impact of adversarial attacks on UAV IDS is largely overlooked in the literature. Our results herein show that evasion attacks pose a significant threat, capable of deteriorating IDS model detection accuracy by 20 %. In this paper, we propose adopting cyber-physical fused datasets to train our proposed unsupervised sequential ensemble learningbased models to improve IDS robustness against evasion attacks. Our results, based on a practical UAV testbed and considering a wide range of evasion attacks, demonstrate that the proposed ensemble of a Transformer autoencoder and long short-term memory recurrent neural network reduces accuracy deterioration to 3 % when combined with the physical and cyber fused features.