PointGuard: Provably Robust 3D Point Cloud Classification

Hongbin Liu, Jinyuan Jia, Neil Zhenqiang Gong

Abstract

3D point cloud classification has many safety-critical applications such as autonomous driving and robotic grasping. However, several studies showed that it is vulnerable to adversarial attacks. In particular, an attacker can make a classifier predict an incorrect label for a 3D point cloud via carefully modifying, adding, and/or deleting a small number of its points. Randomized smoothing is state-of-the-art technique to build certifiably robust 2D image classifiers. However, when applied to 3D point cloud classification, randomized smoothing can only certify robustness against adversarially modified points. In this work, we propose PointGuard, the first defense that has provable robustness guarantees against adversarially modified, added, and/or deleted points. Specifically, given a 3D point cloud and an arbitrary point cloud classifier, our PointGuard first creates multiple subsampled point clouds, each of which contains a random subset of the points in the original point cloud; then our PointGuard predicts the label of the original point cloud as the majority vote among the labels of the subsampled point clouds predicted by the point cloud classifier. Our first major theoretical contribution is that we show PointGuard provably predicts the same label for a 3D point cloud when the number of adversarially modified, added, and/or deleted points is bounded. Our second major theoretical contribution is that we prove the tightness of our derived bound when no assumptions on the point cloud classifier are made. Moreover, we design an efficient algorithm to compute our certified robustness guarantees. We also empirically evaluate PointGuard on ModelNet40 and ScanNet benchmark datasets.

Introduction: The Vulnerability of 3D Point Clouds

3D point clouds—sets of coordinates representing shapes—are the "eyes" of autonomous vehicles and robotic grippers. However, these systems are notoriously fragile. An attacker can subtly modify, add, or delete just a few points to trick a high-performing AI into misidentifying a "plane" as a "radio" or a "tent."

Adversarial attacks on 3D point clouds: modifying, adding, or deleting points can cause misclassification.

While researchers have developed "empirical" defenses to stop these attacks, they often fall into a cat-and-mouse game: a new defense is proposed, and a more clever attack immediately breaks it. To solve this, we need certified robustness—a mathematical guarantee that the model will not change its mind unless the input is altered beyond a certain threshold. While "randomized smoothing" provides this for 2D images by adding noise, it fails for 3D clouds because it cannot account for the structural changes caused by adding or removing points entirely.

Methods: The PointGuard Framework

The authors introduce a framework called PointGuard, which shifts the defense strategy from adding noise to randomized subsampling.

PointGuard defense mechanism: randomized subsampling and majority voting to recover the correct label.
  1. Subsampling: Given a point cloud with $n$ points, PointGuard creates thousands of smaller "subsampled" versions, each containing a random subset of $k$ points.
  2. Majority Voting: A standard 3D classifier (like PointNet) predicts a label for every subsampled version. PointGuard then takes a majority vote to determine the final label.
  3. Certification: The core innovation is a mathematical proof using the Neyman-Pearson Lemma. The authors derive a "certified perturbation size" ($r^*$). If an attacker changes fewer than $r^*$ points, the math guarantees that the majority vote will remain the same.
  4. Robust Training: To make this work, the base classifier is trained specifically on subsampled data, ensuring it stays accurate even when looking at sparse, partial versions of an object.

Datasets and Evaluation

The researchers evaluated PointGuard using two gold-standard benchmarks in the 3D vision community:

  • ModelNet40: A collection of over 12,000 clean 3D CAD models across 40 categories (chairs, planes, etc.).
  • ScanNet: A more challenging dataset containing 3D views of real-world indoor scenes (offices, living rooms) with 16 object categories.

They tested PointGuard on top of two popular architectures: PointNet and DGCNN.

Key Insights

  • Subsampling as the "Noise" of 3D: In 2D images, pixels are fixed in a grid, so we add Gaussian noise. In 3D, the "noise" is the presence or absence of points. Subsampling is the natural 3D analogue to randomized smoothing because it makes the classifier's output dependent on the set of points rather than individual coordinates.
  • The Power of Sparsity: By choosing a small subsample size ($k$), the probability that an attacker's "poisoned" points end up in a specific subsample becomes very low. This allows the "clean" points to dominate the majority vote.
  • Theoretical Tightness: The authors didn't just find a safety bound; they proved their bound is "tight." This means it is mathematically impossible to derive a larger certified robustness size without making specific assumptions about how the underlying classifier works.

Key Results

PointGuard significantly closes the gap between empirical and certified security:

  • Defeating "Total Failure": On ModelNet40, a standard PointNet model drops to 0% accuracy under a 30-point modification attack. Under the same conditions, PointGuard maintains a 69.7% certified accuracy.
  • Outperforming 2D Methods: When compared to traditional randomized smoothing (adding Gaussian noise), PointGuard was substantially more robust against point modification.
  • Versatility: The defense proved effective against all four major attack types: modifying existing points, adding new malicious points, deleting critical points, and small coordinate perturbations.

Limitations and Future Work

  • Computational Overhead: To get a certified prediction, PointGuard must run the classifier thousands of times (e.g., $N=10,000$) for a single object. This makes it difficult to use in real-time applications like high-speed autonomous driving without significant hardware acceleration.
  • The Accuracy-Robustness Trade-off: There is a "sweet spot" for the subsample size $k$. If $k$ is too small, the model becomes very robust but loses "clean" accuracy. If $k$ is too large, it becomes easier to attack.
  • Future Directions: Future work could involve more sophisticated subsampling techniques that preserve the "topology" or essential skeleton of a shape, potentially improving accuracy without sacrificing the safety guarantee.

Summary updated

Abstract

3D point cloud classification has many safety-critical applications such as autonomous driving and robotic grasping. However, several studies showed that it is vulnerable to adversarial attacks. In particular, an attacker can make a classifier predict an incorrect label for a 3D point cloud via carefully modifying, adding, and/or deleting a small number of its points. Randomized smoothing is state-of-the-art technique to build certifiably robust 2D image classifiers. However, when applied to 3D point cloud classification, randomized smoothing can only certify robustness against adversarially modified points. In this work, we propose PointGuard, the first defense that has provable robustness guarantees against adversarially modified, added, and/or deleted points. Specifically, given a 3D point cloud and an arbitrary point cloud classifier, our PointGuard first creates multiple subsampled point clouds, each of which contains a random subset of the points in the original point cloud; then our PointGuard predicts the label of the original point cloud as the majority vote among the labels of the subsampled point clouds predicted by the point cloud classifier. Our first major theoretical contribution is that we show PointGuard provably predicts the same label for a 3D point cloud when the number of adversarially modified, added, and/or deleted points is bounded. Our second major theoretical contribution is that we prove the tightness of our derived bound when no assumptions on the point cloud classifier are made. Moreover, we design an efficient algorithm to compute our certified robustness guarantees. We also empirically evaluate PointGuard on ModelNet40 and ScanNet benchmark datasets.
BibTeX
@inproceedings{liu2021pointguard,
  author = {Hongbin Liu and Jinyuan Jia and Neil Zhenqiang Gong},
  title = {{PointGuard: Provably Robust 3D Point Cloud Classification}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2021},
  url = {https://openreview.net/forum?id=y3-K36wBDd0},
}
Figure