On Adversarial Robustness of Point Cloud Semantic Segmentation

TL;DR

This paper systematically analyzes the adversarial robustness of 3D point cloud semantic segmentation models, revealing their vulnerability and highlighting the effectiveness of point color attacks, thereby urging the development of more robust methods.

Contribution

It provides the first formal analysis of adversarial attacks on PCSS, introduces new attack methods, and evaluates their impact across multiple models and datasets.

Findings

All models are vulnerable to adversarial attacks.

Attacking point color is more effective than other features.

The study emphasizes the need for robust PCSS models.

Abstract

Recent research efforts on 3D point cloud semantic segmentation (PCSS) have achieved outstanding performance by adopting neural networks. However, the robustness of these complex models have not been systematically analyzed. Given that PCSS has been applied in many safety-critical applications like autonomous driving, it is important to fill this knowledge gap, especially, how these models are affected under adversarial samples. As such, we present a comparative study of PCSS robustness. First, we formally define the attacker's objective under performance degradation and object hiding. Then, we develop new attack by whether to bound the norm. We evaluate different attack options on two datasets and three PCSS models. We found all the models are vulnerable and attacking point color is more effective. With this study, we call the attention of the research community to develop new approaches to harden PCSS models.