Machine Learning-Based Intrusion Detection for Swarm of Unmanned Aerial Vehicles

Swarms of unmanned aerial vehicles (UAVs) are widely adopted in civilian and military applications. However, this cyber-physical system is threatened by cyber-attacks. Recently, machine learning-based intrusion detection systems have been successfully adopted to detect cyber-attacks. Yet, the following questions remain unanswered: (a) Can the fusion of cyber and physical features collected from the attacked UAV improve the detection performance? (b) Can the fusion of cyber and physical features collected from unattacked UAVs in the swarm help to detect the attack? (c) Can the fusion of cyber and physical features collected from all UAVs in the swarm (attacked and unattacked) improve the detection performance? To answer the aforementioned questions, and due to the absence of practical datasets, we develop a preliminary testbed of two UAVs flying in coordination. We launch a range of cyber-attacks on one of the UAVs including false data injection (FDI), denial-of-service (DoS), replay, and evil twin attacks. Then, we collect cyber and physical features from the UAVs under normal operation and attack conditions. Next, we develop a set of intrusion detection systems based on shallow and deep machine learning models including support vector machine (SVM), feedforward neural networks (FNN), recurrent neural networks (RNN), and convolutional neural networks (CNN). The developed models are trained using cyber-only, physical-only, and cyber-physical features collected from the attacked UAV, the unattacked UAV, and both UAVs in the swarm. The extensive studies carried out herein provide answers to the aforementioned questions and pave the way toward effective intrusion detection systems in UAV swarms.