Adversarial Attacks on Faster R-CNN Model for Object Detection in Autonomous Vehicles

Object detection models used in autonomous driving systems provide environmental awareness of vehicles by making reliable and accurate predictions. However, these models show significant vulnerabilities in the face of adversarial attacks that can jeopardize system security. In this research, we compare the test performance and evaluation of the Faster R-CNN model under adversarial attacks using the ApolloScape dataset. We apply adversarial attacks to object detection models of autonomous driving systems on the ApolloScape dataset, which contains real-world scenes and high-resolution images for the first time. We analyze the prediction performance of the model using two common attack methods, FGSM and PGD. Our results show that PGD severely reduces the accuracy of the model due to its iterative nature and increases the false detection rate, especially in real-life scenes of the ApolloScape dataset. FGSM, on the other hand, showed a more limited effect, although it caused performance loss in critical categories. This research emphasizes that the Faster R-CNN model, which shows high performance on complex real-world datasets, is vulnerable to adversarial attacks. It also highlights the need for robust defense mechanisms tailored to the challenges of autonomous driving systems. We aim to contribute to the development of safer and more secure object detection systems in the face of hostile threats.