Study of Adversarial Attack Strategies on Autonomous Vehicles equipped with LiDAR Sensors

Researchers will study adversarial strategies targeting autonomous vehicles, examining deep neural network-based 3D object detection models, such as voxel-based, point-net base, and graph neural network-based.

Funded by the CCI Coastal Virginia Node

Project Investigators

• Project Investigator (PI): Abhishek Phadke, Christopher Newport University School of Engineering & Computing
• Co-PI: Pratip Rana, Old Dominion University Department of Computer Science

Rationale

Self-driving cars or autonomous vehicles (AV) equipped with LiDAR (Light Detection and Ranging) sensors are revolutionizing the transportation system. 

Adversarial attacks are a serious threat to the reliability of autonomous vehicles, and  those with LiDAR are particularly vulnerable. 

Fake object injection can trick the system and potentially cause damage, even accidents. The targeted vanishing of LiDAR cloud points is another successful attack strategy, which makes the deep neural networks model misclassify 3D objects.

Projected Outcomes

Researchers will identify defense techniques and identify architectural changes in the deep neural networks model and LiDAR sensors to improve the security and reliability of these autonomous vehicles.

They will create a repository of code developed during the project and share trained attack and defense models on LiDAR systems using publicly available datasets.