NoPain: No-box Point Cloud Attack via Optimal Transport Singular Boundary

Zezeng Li, Xiaoyu Du, Na Lei, Liming Chen, Weimin Wang

Abstract

Adversarial attacks on 3D point clouds have traditionally relied on model-specific gradients or iterative queries, often leading to poor transferability across different architectures. This paper introduces NoPain, a novel "no-box" adversarial attack that shifts the focus from model-centric optimization to data-centric geometric analysis. By leveraging Optimal Transport (OT) theory, NoPain identifies "singular boundaries" inherent in the data distribution's manifold. Perturbing samples toward these boundaries generates highly transferable adversarial examples in a single forward pass, requiring zero knowledge of or queries to the target model. Experimental results on ModelNet40 and ShapeNetPart demonstrate that NoPain achieves superior transferability and orders-of-magnitude faster generation times compared to state-of-the-art methods.

1. Introduction

As 3D point cloud processing becomes integral to safety-critical systems like autonomous driving and robotic perception, the robustness of deep learning models against adversarial attacks has become a primary concern. Existing adversarial strategies generally fall into two categories: white-box attacks, which require full access to model weights, and black-box attacks, which rely on repeated model queries. Both paradigms suffer from significant limitations, most notably a lack of transferability—where an attack designed for one model fails to fool another—and high computational overhead due to iterative optimization.

NoPain (No-box Point Cloud Attack) addresses these challenges by proposing a "no-box" framework. Instead of attacking a specific surrogate model, NoPain targets the underlying data distribution itself. By identifying the geometric "weak spots" or singular boundaries of the data manifold, the method generates adversarial samples that are inherently difficult for any classifier trained on that distribution to categorize correctly.

2. Methodology

NoPain treats adversarial generation as a generative task rather than an optimization problem. The framework consists of four primary stages:

2.1 Latent Embedding

To analyze the data manifold effectively, input point clouds are first mapped into a lower-dimensional latent feature space. This is achieved using a pre-trained autoencoder, which captures the essential geometric features of the 3D objects.

2.2 Optimal Transport and Singular Boundaries

The core innovation of NoPain lies in its use of Optimal Transport (OT). The authors solve a semi-discrete OT problem to map a simple noise distribution to the complex distribution of the latent data features. In OT theory, the mapping between distributions can contain "singular boundaries"—regions where the transport plan is non-differentiable or discontinuous. Geometrically, these boundaries represent the transition zones between different data modes (e.g., the ambiguous space between the manifold of a "chair" and that of a "table").

2.3 Boundary Perturbation

Rather than applying random noise, NoPain calculates the direction toward the nearest singular boundary in the latent space. By shifting the latent features toward these singularities, the method creates "mode-mixed" features. These perturbed features represent points that sit on the edge of the data's natural clusters, making them inherently ambiguous to neural networks.

2.4 Decoding

Finally, the perturbed latent features are passed through a pre-trained decoder. This transforms the mathematical perturbations back into a 3D point cloud. Because the process is a single-step forward pass, it eliminates the need for the hundreds or thousands of iterations required by gradient-based attacks.

3. Experimental Setup

The effectiveness of NoPain was evaluated using two benchmark datasets:

  • ModelNet40: 12,311 CAD models across 40 categories.
  • ShapeNetPart: 16,881 shapes from 16 categories, used for classification and segmentation tasks.

The attack's performance was tested against a diverse set of popular 3D deep learning architectures, including PointNet, PointNet++, DGCNN, PointConv, and PCT.

4. Key Results

4.1 Superior Transferability

NoPain demonstrates remarkable transferability across different architectures. On the ModelNet40 dataset, NoPain achieved an Attack Success Rate (ASR) of up to 100% when transferring adversarial samples between different models. This significantly outperforms existing black-box attacks like AdvPC and SI-Adv, which typically achieve transferability rates between 15% and 50%.

4.2 Computational Efficiency

By bypassing iterative optimization, NoPain is orders of magnitude faster than its predecessors. The Average Generation Time (AGT) for a single adversarial sample is approximately 0.026 seconds. In comparison, AdvPC requires 6.2 seconds, and SS-attack can take over 50 seconds per sample.

4.3 Robustness to Defenses

The adversarial samples generated by NoPain proved resilient against common 3D point cloud defenses. Even when subjected to Statistical Outlier Removal (SOR) and Simple Random Sampling (SRS), NoPain maintained a high ASR, suggesting that the perturbations are deeply embedded in the geometry of the object rather than appearing as simple surface noise.

5. Discussion and Limitations

While NoPain represents a significant advancement in no-box attacks, it possesses certain dependencies. The effectiveness of the attack is intrinsically linked to the quality of the pre-trained autoencoder; if the latent space does not accurately reflect the data manifold, the identified singular boundaries may be less effective. Additionally, the current implementation uses a fixed perturbation budget (threshold $\tau$), which may require further tuning to balance stealthiness (measured by Chamfer Distance) with attack aggressiveness. Future research could also explore the application of OT singular boundaries to complex, multi-object scenes typical of real-world LiDAR data.

6. Conclusion

NoPain demonstrates that the most effective way to attack a neural network may not be to study the network itself, but to study the data it learns from. By identifying and exploiting the singular boundaries of the data manifold via Optimal Transport, NoPain provides a highly efficient, transferable, and robust method for generating adversarial point clouds without requiring any access to the target model's parameters or outputs.

Summary updated

BibTeX
@inproceedings{li2025nopain,
  author = {Zezeng Li and Xiaoyu Du and Na Lei and Liming Chen and Weimin Wang},
  title = {{NoPain: No-box Point Cloud Attack via Optimal Transport Singular Boundary}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.00331},
}
Figure