Researchers Fool LiDAR with 3D-Printed Adversarial Objects
Despite the very high cost of LiDAR, most players in the self-driving technology market regard the advanced spacial surveying technology as an indispensable sensor option for autonomous vehicles. Unlike cameras which can be negatively affected by bad weather or low-light conditions, LiDAR produces accurate, computer-friendly point cloud data across a wide range of conditions and can complement other sensors to improve safety for self-driving vehicles.
However, like most other tech, LiDAR is also vulnerable to hackers. To illustrate this, researchers from Baidu Research, the University of Michigan, and the University of Illinois at Urbana-Champaign have published a method for generating adversarial objects that can befuddle the LiDAR point cloud and compromise security in vehicles using the tech.
The researchers first employed an evolution-based blackbox attack algorithm to demonstrate the vulnerabilities of LiDAR sensors, then applied their new gradient-based approach LiDAR-Adv to explore the effect of powerful adversarial samples. They tested the efficacy of LiDAR-Adv on Baidu’s home-grown Apollo autonomous driving platform and in the real world. What’s scary is that LiDAR-Adv did not only fool LiDAR in simulated environments, 3D-printed adversarial samples also went undetected by LiDAR in the real world.
LiDAR-based detection systems consist of multiple non-differentiable steps rather than a single end-to-end network. This significantly limits the effectiveness of gradient-based end-to-end attacks and had so far helped protect LiDAR from adversarial attacks. To create effective adversarial samples, researchers faced three challenges :
- A LiDAR-based detection system uses a solid LiDAR device to map a 3D shape onto a point cloud, which is then fed into a machine learning detection system. Therefore, how shape perturbation could affect the scanned point cloud is unclear.
- Since the gradient-based optimizer does not work on preprocessing of the LiDAR point cloud, a new optimizer is needed.
- The perturbation space is subject to various restrictions.
To meet these challenges researchers first simulated a differentiable LiDAR renderer that could link perturbations of 3D targets to a LiDAR scan (or point cloud). Then they made 3D feature aggregations using the differentiated proxy function. Finally, they designed different losses to ensure that the generated 3D adversarial samples were smooth. The results show that with 3D sensing and product-level multi-stage detectors, researchers are able to mislead an autonomous driving system.
To better demonstrate the flexibility of LiDAR-Adv attack methods, researchers used two additional evaluation scenarios:
- Hidden target: synthesize an adversarial sample that will not be detected;
- Change Label: synthesize an adversarial sample that is recognized as a specific target.
In the experiments researchers succeeded in creating undetectable targets, exposing vulnerabilities in LiDAR detection systems through an evolution-based black box algorithm. Researchers then also demonstrated the qualitative and quantitative results of the LiDAR-Adv method in a white box setting. Moreover, the experiment results showed that the LiDAR-Adv method can also achieve other malicious and potentially dangerous targets such as “Change Label”.
The paper Adversarial Objects Against LiDAR-Based Autonomous Driving Systems is on arXiv. Further information on the experiments can be found here.
Author: Reina Qi Wan | Editor: Michael Sarazen
2018 Fortune Global 500 Public Company AI Adaptivity Report is out!
Purchase a Kindle-formatted report on Amazon.
Apply for Insight Partner Program to get a complimentary full PDF report.
Follow us on Twitter @Synced_Global or daily AI news!
We know you don’t want to miss any stories. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.