Understanding and implementing defenses for specific Adversarial Attacks on the YOLOP model
Autonomous vehicles (AVs) leverage machine-learning perception models to detect and classify critical objects such as road signs, vehicles, lane lines, hazards, and pedestrians, enabling self-driving functionalities. With the nationwide proliferation of AVs, the demand for safe, secure, accurate, and rapid driving perception models has surged dramatically. Panoptic perception models have been proposed to offer advanced object detection and segmentation capabilities for AVs. This work explores the robustness of panoptic perception models against adversarial attacks, focusing on the YOLO-P (You Only Look Once for Panoptic Driving Perception) model. To comprehensively evaluate the safety of panoptic perception models, we subject the model to various adversarial attacks, including the Fast Gradient Sign Method (FGSM), Jacobian-based Saliency Map Attack (JSMA), Color Channel Perturbations (CCP), and Universal Adversarial Perturbations (UAP), to assess their effects on model performance. Subsequent defenses, including image pre-processing techniques and the deployment of a Defense Generative Adversarial Network (GAN), are implemented to mitigate attack effects. Our findings reveal deprecated performance post-attack implementation, with only marginal improvement in the performance of models post-defense application. These results suggest the necessity for further research into more effective defense mechanisms to ensure the safety and reliability of AVs against adversarial threats.
Acknowledgments
This research work was conducted at Oakland University in the REU program and partially supported by the NSF REU grant CNS-2349663.