Abstract
NoPain is a novel "no-box" adversarial attack for 3D point clouds that requires zero knowledge of the target model or its gradients. By leveraging Optimal Transport (OT) to identify inherent "singular boundaries" within the data distribution, NoPain generates highly transferable adversarial samples that effectively fool diverse architectures with near-instantaneous generation times. Unlike traditional methods that target classifier decision boundaries, NoPain exploits the intrinsic geometric vulnerabilities of the data manifold.
Introduction
Adversarial attacks on 3D point clouds are critical for evaluating the robustness of systems in safety-critical applications like autonomous driving. Existing attacks typically fall into two categories: white-box (full model access) and black-box (querying a surrogate model). A major bottleneck for these methods is transferability; attacks optimized against one specific model often fail when tested against different architectures. Furthermore, these attacks usually require iterative optimization, which is computationally expensive.
NoPain addresses these challenges by shifting the focus from the classifier's decision boundaries to the intrinsic geometric boundaries of the data manifold. By attacking the data distribution itself, the framework creates adversarial samples that are naturally effective across any model trained on that data.
Methodology
The NoPain framework operates in three main stages without requiring access to a classifier or its gradients:
- Latent Embedding: The input point cloud is compressed into a latent feature space $Y$ using a pre-trained autoencoder (e.g., PointFlow or Point-Diffusion).
- Optimal Transport (OT) Mapping: The core innovation involves solving a semi-discrete OT problem to map a continuous noise space $X$ to the discrete feature space of the data. This mapping induces a "cell decomposition" of the space into hyperplanes.
- Singular Boundary Determination: According to OT theory, "singular boundaries" emerge in regions where the transport mapping is non-differentiable. The authors identify these by calculating the dihedral angles between neighboring hyperplanes. Large angles indicate "seams" or "tears" in the data manifold.
- No-Box Attack: Instead of iterative optimization, NoPain simply samples new points along these identified singular boundaries. These "mode-mixed" samples are then decoded back into 3D point clouds. Because these boundaries are inherent to the data, they represent universal "weak spots."
Experimental Results
The authors evaluated NoPain on two standard 3D benchmarks: ModelNet40 (12,311 CAD models) and ShapeNetPart (16,881 shapes).
Superior Transferability and Efficiency
In tests on ModelNet40, the diffusion-based version (NoPain-PD) achieved a 100% Attack Success Rate (ASR) against PointNet and over 82% ASR against DGCNN and PCT, significantly outperforming existing black-box transfer attacks like AdvPC and SI-Adv.
Crucially, NoPain is a "one-shot" generative process. While traditional attacks take several seconds or minutes to generate a single sample, NoPain generates samples in roughly 0.02 to 0.03 seconds—a speedup of several orders of magnitude measured by the Average Time Cost (ATC) metric.
Defense Robustness
The attack remained effective even against specialized 3D defenses like SRS (Statistical Outlier Removal) and DUP-Net. This suggests that the adversarial features generated are "intrinsic" to the data geometry rather than simple high-frequency noise that can be easily filtered.
Conclusion
NoPain demonstrates that adversarial vulnerability is not just a flaw in neural network training but is often baked into the geometry of the data distribution itself. By using dihedral angles between OT hyperplanes, the authors provide a geometric interpretation of model confusion. While the method currently relies on the quality of the latent embedding from a pre-trained autoencoder, it represents a significant step toward efficient, model-agnostic adversarial research in 3D computer vision.
Summary updated