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
3D point clouds—sets of coordinates representing shapes—serve as the primary sensory input for autonomous vehicles and robotic systems. Despite their widespread use, these systems are notoriously fragile. Adversarial attacks can trick state-of-the-art classifiers into misidentifying objects (e.g., a "plane" as a "radio") by subtly moving, adding, or deleting just a few points.
While "randomized smoothing" (adding Gaussian noise) provides robustness for 2D image classifiers, it is less effective for 3D data. Standard smoothing techniques typically only certify against continuous coordinate shifts (perturbations) and fail to handle structural attacks like point addition or removal, which alter the input size. PointGuard addresses this gap as the first defense providing provable robustness guarantees against point modification, addition, and deletion.
Methodology: Randomized Subsampling
The core of PointGuard is a "subsample-and-vote" strategy that treats point cloud classification as a discrete problem. The process involves three main steps:
- Subsampling: For a given point cloud, PointGuard generates multiple subsampled versions, each containing a random subset of $k$ points.
- Majority Voting: A base classifier (e.g., PointNet) predicts a label for each subsampled cloud. The final output is the label with the majority of votes.
- Certification: Using the Neyman-Pearson Lemma, the authors derive a certified perturbation size $r^*$. If an attacker alters fewer than $r^*$ points, the majority vote is mathematically guaranteed to remain unchanged.
To optimize performance, the authors introduce subsample-aware training. Standard models trained on full point clouds often struggle with the sparsity of subsampled data. Training the base classifier specifically on these smaller subsets significantly improves both clean accuracy and certified robustness.
Experimental Evaluation
PointGuard was evaluated on two benchmark datasets: ModelNet40 (12,311 CAD models across 40 categories) and ScanNet (indoor scenes with 16 object categories). The authors proved that their derived robustness bounds are tight, meaning that without specific assumptions about the underlying classifier, it is mathematically impossible to guarantee robustness for a larger number of points.
The Robustness Trade-off
The subsample size $k$ is a critical hyperparameter. Smaller $k$ values increase robustness—making it harder for an attacker to "infect" enough subsamples to flip the vote—but can reduce accuracy on clean data due to the loss of information.
Results and Discussion
PointGuard demonstrates significant improvements over existing methods and undefended models. For point modification attacks on ModelNet40, it substantially outperforms the 3D adaptation of randomized smoothing.
As shown in the results, a standard PointNet model's accuracy drops to 0% when just 30 points are modified. In contrast, PointGuard (with $k=16$) maintains a certified accuracy of 69.7% under the same conditions. Furthermore, PointGuard provides the first-ever certified guarantees against point addition and deletion, maintaining robustness even when hundreds of points are removed.
While PointGuard introduces some computational overhead due to the multiple forward passes required for voting, it represents a major step forward in securing 3D perception systems for safety-critical applications.
Summary updated