Robust Unsupervised Domain Adaptation for 3D Point Cloud Segmentation under Source Adversarial Attacks
Haosheng Li, Junjie Chen, Yuecong Xu, KEMI DING
AI summary
Problem
Existing unsupervised domain adaptation methods for 3D point clouds ignore vulnerabilities to subtle adversarial perturbations and noisy labels in the source domain, leading to severe performance degradation in real-world deployments.
Approach
The authors introduce the AdvSynLiDAR dataset with perceptually-aware attacks and propose the Adversarial Adaptation Framework (AAF), which combines a Robust Long-Tailed Loss for pre-training with a probability decoder and high-confidence neighborhood aggregation during adaptation to restore geometry and generate reliable pseudo-labels.
Key results
- First exploration of adversarial robustness in 3D point cloud cross-domain adaptation
- Introduction of AdvSynLiDAR dataset simulating stealthy source-domain perturbations and label noise
- Development of AAF framework with Robust Long-Tailed Loss and probabilistic decoder for robust adaptation
- mIoU improvements of 11.61 on SemanticKITTI and 9.85 on SemanticPOSS under adversarial conditions
Why it matters
Enables reliable deployment of 3D perception models in autonomous driving and robotics where source data integrity cannot be guaranteed.
Abstract
Unsupervised domain adaptation (UDA) frame- works have shown good generalization capabilities for 3D point cloud semantic segmentation models on clean data. However, existing works overlook adversarial robustness when the source domain itself is compromised. To comprehensively explore the robustness of the UDA frameworks, we first design a stealthy adversarial point cloud generation attack that can significantly contaminate datasets with only minor perturbations to the point cloud surface. Based on that, we propose a novel dataset, Ad- vSynLiDAR, comprising synthesized contaminated LiDAR point clouds. With the generated corrupted data, we further develop the Adversarial Adaptation Framework as the countermeasure. Specifically, by extending the key point sensitive loss towards the Robust Long-Tailed loss and utilizing a decoder branch, our approach enables the model to focus on long-tailed classes during the pre-training phase and leverages high-confidence decoded point cloud information to restore point cloud structures during the adaptation phase. We evaluated our method on the AdvSyn- LiDAR dataset, where the results demonstrate that our method can mitigate performance degradation under source adversarial perturbations for UDA in the 3D point cloud segmentation application.