Novelty-based Intrusion Detection of Sensor Attacks on Unmanned Aerial Vehicles

Unmanned Aerial Vehicles (UAVs) have proven to be a useful technology in numerous industries including industrial control systems surveillance, law enforcement, and military operations. Due to their heavy reliance on wireless protocols and hostile operating environments, UAVs face a large threat landscape. As attacks against UAVs increase, an intelligent intrusion detection system (IDS) is needed to aid the UAV in identifying attacks. The UAV domain presents unique challenges for intelligent IDS development, such as the variety of sensors, communication protocols, UAV platforms, control configurations, and dataset availability. In this paper, we propose a novelty-based approach to intrusion detection in UAVs by using one-class classifiers. One-class classifiers require only non-anomalous data to exist in the training set. This allows for the use of flight logs as training data, which are created by most UAVs during flight by default. Principal Component Analysis is applied to sensor logs for dimensionality reduction, and one-class classifier models are generated per sensor. A number of one-class classifiers are selected: One-Class Support Vector Machine, Autoencoder Neural Network, and Local Outlier Factor. The pre-processing, feature selection, training, and tuning of the selected algorithms is discussed. GPS spoofing is used throughout the paper as a common example of an external sensor-based attack. This approach shows to be effective across multiple UAV platforms with platform-specific F1 scores up to 99.56% and 99.73% for benign and malicious sensor readings respectively.