TL;DR
This paper introduces GeoA³, a novel method for generating adversarial point clouds that are imperceptible to humans and harder to defend against. By incorporating geometry-aware objectives that promote surface smoothness and fairness, GeoA³ produces more realistic adversarial examples compared to existing methods, which often result in noticeable outliers.
What Problem Does This Paper Address?
Machine learning models, particularly those used for 3D point cloud classification, are vulnerable to adversarial examples – subtly modified inputs that cause misclassification. While adversarial attacks on 2D images have been extensively studied, extending these to 3D point clouds presents unique challenges. Existing methods for generating adversarial point clouds often produce noticeable outliers, making them easy to detect and remove. This is because human perception of 3D shapes differs from 2D images; while 2D perception can overlook high-frequency noise, 3D perception is sensitive to surface irregularities like "spiky" points. The paper aims to bridge this gap by creating adversarial point clouds that are both effective at fooling models and imperceptible to human observers, mimicking how humans perceive 3D shapes.
Methods
The core of the proposed method, Geometry-Aware Adversarial Attack (GeoA³), lies in its novel geometry-aware objectives. These objectives are designed to ensure that the generated adversarial point clouds exhibit desirable surface properties like smoothness and fairness, similar to benign point clouds. Specifically, GeoA³ combines:
- Chamfer Distance: A standard metric for comparing point sets, measuring the average distance between points in one set and their nearest neighbors in the other.
- Hausdorff Distance: A more sensitive metric that finds the maximum of the minimum distances between points, helping to penalize outliers.
- Consistency of Local Curvatures: A new term that encourages the local surface geometry (specifically, curvature) of the adversarial point cloud to be consistent with the original point cloud. This is achieved by comparing discrete curvature measures derived from local neighborhoods of points.
These geometry-aware regularizers are used in conjunction with a targeted misclassification loss function, which is designed to achieve higher maliciousness than traditional margin-based losses.
For generating adversarial surface shapes (rather than just point clouds), the paper introduces Geo+A³-IterNormPro. This method builds upon GeoA³ by adding a uniformity-promoting objective and an Iterative Normal Projection algorithm. The uniformity objective prevents adversarial effects from arising solely from irregular point distribution on the surface. The iterative projection algorithm deforms the underlying surface more explicitly, aiming for adversarial effects that persist even after re-sampling from reconstructed meshes.
Datasets
The paper primarily uses the ModelNet40 dataset for training and evaluation. This dataset consists of 12,311 CAD models across 40 semantic categories. For point clouds, they follow a standard procedure of sampling points uniformly from these CAD models.
Key Insights and Ideas
- Human Perception of 3D Shapes is Different: The paper highlights a crucial difference in how humans perceive 2D images versus 3D shapes. While 2D perception can tolerate high-frequency noise, 3D perception is highly sensitive to surface irregularities, making outlier-based adversarial attacks easily detectable.
- Geometry Matters for Imperceptibility: Adversarial point clouds that maintain surface properties like smoothness and fairness are more likely to be imperceptible to humans. This is the central idea behind the proposed geometry-aware objectives.
- Outliers are the Achilles' Heel: Existing adversarial point cloud methods often generate outliers, which are both visually noticeable and easily defended against using simple outlier removal techniques. The proposed method directly addresses this by penalizing such irregularities.
- Surface-Level Attacks are More Robust: Generating adversarial effects that persist even after reconstructing and re-sampling from the surface (surface-level attacks) is a more challenging but also more robust form of adversarial attack, as it's harder to defend against by simple re-sampling.
Key Results
- Improved Attack Success Rate: GeoA³ consistently achieves higher attack success rates compared to existing methods like Xiang et al. [11], Liu et al. [21], and Tsai et al. [34], even after applying state-of-the-art defenses like Statistical Outlier Removal (SOR).
- Enhanced Geometric Regularity: The proposed method significantly improves geometric regularity (measured by R(P')), producing point clouds with smoother surfaces and fewer noticeable outliers.
- Subjective User Studies: Human evaluations on Amazon Mechanical Turk showed that participants preferred GeoA³'s adversarial examples as being visually more similar to the benign ones, indicating superior imperceptibility.
- Surface-Level Attack Efficacy: The Geo+A³-IterNormPro method demonstrated promising results in surface-level adversarial attacks, with a notable percentage of attacks remaining successful even after 3D printing and re-scanning.
Limitations and Future Work
- Surface-Level Attack Challenge: While Geo+A³-IterNormPro shows promise, surface-level adversarial attacks remain a challenging task, with success rates still lower than point cloud attacks.
- Future Directions: The authors suggest future research could focus on improving attacking performance through advanced methods for adversarial point cloud generation or by directly learning to generate adversarial meshes. They also plan to explore more sophisticated geometry-aware objectives and algorithms.
Summary updated