Trojan Attack and Defense for Deep Learning-Based Navigation Systems of Unmanned Aerial Vehicles

As unmanned aerial vehicles (UAVs) become increasingly integrated across various domains, both military and civilian, safeguarding the security of their navigation systems becomes paramount. In the contemporary age, the prominence of cybersecurity for UAVs has grown due to a rising number of cyberattacks on these systems. Notably, over the past decade, several significant cybersecurity breaches have impacted UAVs due to inadequate vulnerability assessments and security measures. Deep learning (DL)-based algorithms show immense potential for enabling autonomous UAV navigation. However, these algorithms are susceptible to malicious attacks, such as DL-based Trojan attacks, which can compromise the integrity and reliability of UAV navigation systems. This paper addresses potential vulnerabilities in DL-based UAV navigation systems and emphasizes the importance of securing these systems against DL-based Trojan attacks. We design various trigger patterns for collision and steering angle of the DroNet model incorporating adversarial inputs to test the robustness of the deep learning algorithm used for UAV navigation. By simulating potential attacks and studying their effects, we aim to highlight the weaknesses and potential entry points for malicious interference. We assess the effectiveness of Trojan attacks on the DroNet model using poisoned collision and steering angle datasets. Subsequently, we regulate the intensity of the designed triggers and evaluate the performance of the DroNet architecture. Additionally, we propose mitigation strategies to enhance the robustness and security of navigation systems against these attacks. To identify the likelihood of the trained model being trojaned or not, we have developed a Trojan detector and created distinct DroNet Trojan Model Datasets for this purpose. That the DroNet model is vulnerable to DL-based Trojan attacks, as evidenced by the successful manipulation of collision and steering angle predictions. The Trojan detector effectively identifies potential compromises within the model, highlighting the necessity for enhanced security measures.