Extending Adversarial Attacks and Defenses to Deep 3D Point Cloud Classifiers

Daniel Liu, Ronald Yu, Hao Su

TL;DR

This paper explores the vulnerability of 3D point cloud classifiers (like PointNet and PointNet++) to adversarial attacks—small, intentional perturbations designed to fool AI. The authors demonstrate that while these 3D models are highly susceptible to adapted 2D attack techniques, they are also surprisingly easy to defend using simple geometric filters that exploit the unique structure of 3D data.

What Problem Does This Paper Address?

As deep learning moves into safety-critical 3D applications—such as autonomous driving and robotics—the robustness of these systems becomes paramount. In the 2D image world, "adversarial examples" (images with imperceptible noise that cause misclassification) are a well-known threat.

However, 3D point clouds are fundamentally different from 2D images. Images are dense, regular grids of pixels, whereas point clouds are sparse, unordered sets of coordinates in 3D space. This paper investigates whether the adversarial "arms race" translates to 3D: Can we fool 3D classifiers with tiny point shifts, and can we build defenses that specifically leverage 3D geometry?

Methods

The authors evaluate several attack and defense strategies by adapting established 2D methods to the 3D domain:

Attacks

  • Gradient-Based Attacks: They adapted the Fast Gradient Sign Method (FGSM) and its iterative versions to shift 3D coordinates. They introduced a "Normalized Gradient $L_2$ Method" to ensure perturbations are distributed more naturally across 3D space.
  • Reducing Perceptibility: To make attacks harder for humans to spot, they proposed:
  • Gradient Projection: Projecting perturbed points back onto the original object's surface (mesh), effectively changing the density of points rather than the shape.
  • Clipping Norms: Limiting point shifts based on the average distance to neighboring points to prevent "floating" outliers.
  • JSMA (Jacobian-based Saliency Map Attack): A targeted attack that modifies only a small subset of the most "influential" points.

Defenses

  • Adversarial Training: Training the model on a mix of clean and adversarial examples.
  • Outlier Removal: A simple geometric defense that calculates the distance of each point to its $k$-nearest neighbors and removes points that are statistically distant from the rest.
  • Saliency Removal: Discarding points that have the highest impact on the model’s gradient.

Datasets

  • ModelNet-40: A standard benchmark containing 3D CAD models from 40 categories (chairs, airplanes, etc.).
  • ModelNet-Unique: A subset of 16 hand-picked classes from ModelNet-40 created by the authors. These classes have highly distinct shapes (e.g., "person" vs. "car") to better measure the effectiveness of attacks that force a model to jump between very different categories.

Key Insights and Ideas

  • 3D is "Easier" to Defend: Unlike 2D images, where adversarial noise is often hidden in high-frequency pixel variations, 3D attacks often create "outliers"—points that float away from the main object. Because physical objects are generally continuous surfaces, these outliers are mathematically easy to detect and prune.
  • The Max-Pooling "Shield": Architectures like PointNet use a global max-pooling layer. This means the model only "sees" a subset of the points (the "critical points"). Attacks are ineffective if they perturb points that don't survive the max-pooling operation, providing a level of inherent structural resistance.
  • Complexity $\neq$ Robustness: Interestingly, PointNet++ (a more complex, hierarchical version of PointNet) was found to be more vulnerable to iterative gradient attacks than its simpler predecessor, though it was more resistant to saliency-based attacks (JSMA).

Key Results

  • Attack Success: On undefended PointNet models, iterative $L_2$ attacks achieved a 90.1% success rate in misclassifying objects.
  • Defense Effectiveness: Simple Outlier Removal proved to be the most robust defense. It reduced the success rate of the JSMA attack from 40.9% to just 0.5%, while only causing a minor 2.6% drop in accuracy on clean data.
  • Transferability: The authors found that adversarial examples generated for one model could often fool another (black-box attack), though success rates were lower (around 10-30%) compared to white-box attacks.

Limitations and Future Work

  • Assumption of Uniform Sampling: The outlier removal defense assumes that clean 3D data is somewhat uniformly sampled. In real-world LiDAR data (like from a self-driving car), sensor noise or occlusions might create "natural" outliers that confuse this defense.
  • Mesh Dependency: The "Gradient Projection" attack requires access to the underlying triangular mesh of the object, which isn't always available in raw point cloud pipelines.
  • Future Directions: The authors suggest exploring more sophisticated "surface-aware" attacks that can bypass simple outlier filters by keeping perturbations strictly within the manifold of the object.

Summary updated

Abstract

3D object classification using deep neural networks has been extremely successful. As the problem of identifying 3D objects has many safety-critical applications, the neural networks have to be robust against adversarial changes to the input data set. We present a preliminary evaluation of adversarial attacks on 3D point cloud classifiers by evaluating adversarial attacks that were proposed for 2D images, and extending those attacks to reduce the perceptibility of the perturbations in 3D space. We also show the effectiveness of simple defenses against those attacks. Finally, we attempt to explain the effectiveness of the defenses through the intrinsic structures of both the point clouds and the neural networks. Overall, we find that 3D point cloud classifiers are weak to adversarial attacks, but they are also more easily defensible compared to 2D image classifiers. Our investigation will provide the groundwork for future studies on improving the robustness of deep neural networks that handle 3D data.
BibTeX
@inproceedings{liu2019extending,
  author = {Daniel Liu and Ronald Yu and Hao Su},
  title = {{Extending Adversarial Attacks and Defenses to Deep 3D Point Cloud Classifiers}},
  booktitle = {ICIP},
  year = {2019},
  url = {https://openreview.net/forum?id=xKCEmqhe9a},
}

Loading related content...

Figure