Explainability-Aware One Point Attack for Point Cloud Neural Networks

Hanxiao Tan, Helena KotthausWACV 2023

TL;DR

This paper introduces novel adversarial attack methods, One Point Attack (OPA) and Critical Traversal Attack (CTA), that leverage explainability techniques to target crucial points in 3D point cloud data. These attacks achieve high success rates with minimal perturbations, offering a new avenue for understanding the vulnerabilities and robustness of point cloud neural networks.

What Problem Does This Paper Address?

Point cloud (PC) neural networks are increasingly vital for applications like autonomous driving and robotics. However, their robustness and reliability are not as well-studied as their 2D image counterparts, posing potential safety risks. While adversarial attacks have been explored, existing methods often focus on deceiving human perception or use less refined methods for identifying critical points. This paper addresses the need for more precise and explainability-driven adversarial attacks on PC networks, aiming to uncover vulnerabilities by targeting the most influential points for a given prediction, rather than just minimizing overall perturbation.

Methods

The core of this work lies in two novel adversarial attack methods:

  • One Point Attack (OPA): This is an extreme form of attack that aims to fool the network by shifting only one point. It leverages explainability methods, specifically Integrated Gradients (IG), to identify the single most critical point for a given prediction. The gradients are then masked to only allow perturbations on this single point, and an optimization process (using Adam) is employed to shift it.
  • Critical Traversal Attack (CTA): Recognizing that the optimal number of perturbed points can vary per instance, CTA adaptively searches for the minimum number of points required to successfully attack a PC instance. It starts by attempting to attack with one point and incrementally increases the number of perturbed points until the attack is successful or a failure condition is met. Like OPA, CTA uses IG to identify critical points.

Both methods incorporate a loss function that balances fooling the classifier (maximizing the activation of the target neuron) and maintaining imperceptibility (minimizing the distance between the original and adversarial point cloud). They also explore different distance metrics like Chamfer and Hausdorff distances for measuring imperceptibility.

Datasets

The primary dataset used for evaluation is ModelNet40, a widely used benchmark for point cloud classification containing 12,311 CAD models across 40 categories. The authors also conducted experiments on the ShapeNet dataset.

Key Insights and Ideas

  • Explainability as a Vulnerability Tool: The paper demonstrates that by using explainability methods (like Integrated Gradients) to identify "critical points," adversarial attacks can be made significantly more effective and targeted. This shifts the focus from generic perturbations to understanding which specific input features are most influential.
  • The Power of a Single Point: The success of OPA, which perturbs only one point, highlights the extreme sensitivity of some PC networks to even minor changes in highly influential input features. This suggests that understanding these critical points is key to understanding network vulnerabilities.
  • Structural Stability Matters: The research reveals that different pooling architectures within PC networks have a significant impact on their robustness. Max-pooling, which concentrates attribution on a few points, leads to less robust networks compared to average or sum-pooling, which distribute attribution more broadly. Median-pooling, however, shows remarkable immunity, likely due to its insensitivity to extreme values.
  • Counterfactual Explanations for PCs: The work positions adversarial attacks as a tool for generating counterfactual explanations for point cloud models, making them more interpretable by showing minimal input changes that alter predictions.

Key Results

  • High Success Rates with Minimal Perturbations: OPA achieved a success rate of 98.7% on ModelNet40 with only one point shifted. CTA achieved 100% success rate with an average of 2 points shifted (for $\beta=0$).
  • Superior Imperceptibility: Compared to other point-shifting attacks, OPA and CTA achieve significantly higher sparsity in the perturbation matrix, meaning fewer points are altered.
  • Transferability: Adversarial examples generated by attacking PointNet showed some transferability to other networks like PointNet++, DGCNN, and PointMLP, though the latter three demonstrated higher stability.
  • Targeted Attacks: CTA proved effective even in targeted attack scenarios, achieving high success rates with minimal perturbations, outperforming existing methods like LG-GAN in terms of perturbation magnitude.
  • Impact of Pooling: OPA's success rate dropped dramatically from 98.7% to 44.8% (average-pooling) and 16.7% (sum-pooling) when max-pooling was replaced, indicating enhanced stabilization.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that OPA's performance can be deteriorated when converting to targeted attacks due to its rigid restriction on the number of perturbed points.
  • While IG is used, the paper notes that other gradient-based explainability methods (Vanilla Gradients, Guided Back-propagation) were also tested, with IG showing superior performance in identifying critical points.
  • Future Work:
    • The authors suggest filtering out critical points in advance as a promising direction to improve explainability.
  • They express interest in developing higher-quality and more human-understandable explanations for PC networks.
  • Further investigation into the robustness of different pooling architectures and the role of attribution distribution is suggested.

Summary updated

BibTeX
@inproceedings{tan2023explainability,
  author = {Hanxiao Tan and Helena Kotthaus},
  title = {{Explainability-Aware One Point Attack for Point Cloud Neural Networks}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2023},
}
Figure