Efficient Detection of GPS Spoofing Attacks on Unmanned Aerial Vehicles Using Deep Learning

Unmanned Aerial Vehicles (UAV) have gained major attention in recent years due to its numerous benefits. The navigation of UAVs is heavily reliant on sensors such as the Global Positioning System (GPS). However, sensor attacks that specifically target the GPS are a major concern. Equipment such as the software-defined radio can be used to launch GPS spoofing attacks. Existing techniques used to monitor the GPS and channel characteristics as a way to detect these attacks are insufficient due to the clear knowledge of the structure of GPS and channel characteristics which gives adversaries a blueprint to launch stealthy attacks. In this work, we developed various deep learning models that rely on UAV flight logs and telemetry data to detect GPS spoofing attacks in real time. We generate UAV data for different UAV models and use the data to train the models to detect the GPS spoofing attacks through classification. We developed two types of detectors: a Long Short-Term Memory (LSTM) binary classifier and an LSTM autoencoder-based one-class classifier (OCC). For each type, we compared two variations namely a UAV-specific detector that is customized for different UAV types and a UAV-generalized detector that can work with any model of UAV. We test and compare the efficiency of detection models in simulation. Hardware validation experiments were also performed using Intel® NCS2. For the binary classifier, we get a detection accuracy of 97.79% for the UAV-generalized detector and up to 99.56% for the UAV-model-specific detector. While for the OCC, we get 94.98% for the UAV-generalized detector and up to 99.24% for the UAV-model-specific detector.