Adversarial Attacks on Deep Learning-Based UAV Navigation Systems

In recent years, the prevalence of unmanned aerial vehicles (UAVs) has grown significantly in both military and civilian settings due to their versatile capabilities and the ability to undertake various tasks. Nonetheless, the rise of UAVs has also brought about a significant concern - the cybersecurity aspect. The increasing number of cyberattacks targeting drone systems has become a major issue. Unfortunately, inadequate vulnerability assessments and security countermeasures have resulted in numerous serious cybersecurity breaches on UAVs, highlighting the urgent need for enhanced protection measures. In this paper, we investigate the adversarial attacks on Deep Learning-based UAVs and present five methods for carrying out adversarial attacks on DroNet model used in UAVs. DroNet is a convolutional neural network that guides UAVs by predicting steering angles and collision probabilities based on camera images. We created several adversarial input images for the collision data and steering angle data for attacking purposes. By perturbing a small number of samples from the clean dataset, resulting in a significant reduction in model performance by 18.66% for the collision dataset, and examining the root mean square error rate for the steering angle dataset, which increased the error rate by 427.52%, we observed the impact of sample perturbation on model performance and error rates.