TPC: Transformation-Specific Smoothing for Point Cloud Models

Wenda Chu, Linyi Li, Bo Li

Abstract

Point cloud models with neural network architectures have achieved great success and have been widely used in safety-critical applications, such as Lidar-based recognition systems in autonomous vehicles. However, such models are shown vulnerable to adversarial attacks which aim to apply stealthy semantic transformations such as rotation and tapering to mislead model predictions. In this paper, we propose a transformation-specific smoothing framework TPC, which provides tight and scalable robustness guarantees for point cloud models against semantic transformation attacks. We first categorize common 3D transformations into three categories: additive (e.g., shearing), composable (e.g., rotation), and indirectly composable (e.g., tapering), and we present generic robustness certification strategies for all categories respectively. We then specify unique certification protocols for a range of specific semantic transformations and their compositions. Extensive experiments on several common 3D transformations show that TPC significantly outperforms the state of the art. For example, our framework boosts the certified accuracy against twisting transformation along z-axis (within 20$^\circ$) from 20.3$\%$ to 83.8$\%$.

Introduction

Point clouds serve as the primary sensory input for many autonomous systems, particularly those utilizing LiDAR. While deep learning models excel at identifying objects in these 3D datasets, they are notoriously brittle. An attacker can often mislead a model by applying simple semantic transformations—such as rotating an object, shearing its coordinates, or "tapering" its shape—to trick a model into misclassifying a car as a table.

Overview of the TPC framework showing how it provides certified robustness against semantic transformations like rotation.

The core challenge lies in Certified Robustness. Beyond empirical performance, safety-critical applications require mathematical guarantees that a model's prediction remains constant within a specific range of transformations (e.g., ensuring a car is classified as such for any rotation within ±20°). Previous certification methods often provided "loose" guarantees or failed to scale as point cloud density increased.

The TPC Framework

The proposed TPC (Transformation-Specific Smoothing) framework addresses these challenges by categorizing common 3D transformations into three distinct mathematical classes, allowing for tailored certification strategies:

  1. Additive Transformations: (e.g., shearing or twisting along an axis). These are mathematically well-behaved and certified using randomized smoothing with Gaussian noise.
  2. Composable Transformations: (e.g., general 3D rotations). These allow two sequential steps to be combined into a single step of the same type.
  3. Indirectly Composable Transformations: (e.g., tapering). These do not follow simple composition rules and require more complex handling.
Categorization of 3D transformations into Additive, Composable, and Indirectly Composable classes.

For complex categories, TPC employs a sampling-based certification strategy. Rather than certifying the entire transformation space at once, the framework samples the parameter space and bounds the interpolation error between samples. This creates a continuous "safety zone" of guaranteed model stability.

Experimental Evaluation

TPC was evaluated across standard benchmarks including ModelNet40 for object classification and ShapeNet for part segmentation. The framework was tested on architectures like PointNet and CurveNet, demonstrating its model-agnostic nature.

Key Results

TPC demonstrates dramatic improvements over the previous state-of-the-art, DeepG3D:

  • Twisting Attacks: Against a ±20° twist along the z-axis, TPC achieved 83.8% certified accuracy, compared to 20.3% for the baseline.
  • Shearing: Against a 0.03 radius shear, TPC reached 83.4% accuracy (vs. 59.8%).
  • Scalability: On a 1024-point model, while the baseline's certification for a 3° rotation dropped to 32.3%, TPC maintained 90.5%.

Unlike previous "relaxation-based" methods where guarantees weaken with more points, TPC benefits from higher point density; as the point cloud grows, the certified accuracy increases because the underlying model becomes more confident.

Conclusion and Future Work

TPC provides the first tight, scalable, and mathematically certified robustness guarantees for point cloud models against semantic attacks. While a major advancement, the framework currently faces higher computational costs due to the Monte Carlo sampling required for certification. Future work may focus on tightening bounds for $L_\infty$ coordinate noise and extending the framework to handle arbitrary, non-rigid deformations.

Summary updated

Abstract

Point cloud models with neural network architectures have achieved great success and have been widely used in safety-critical applications, such as Lidar-based recognition systems in autonomous vehicles. However, such models are shown vulnerable to adversarial attacks which aim to apply stealthy semantic transformations such as rotation and tapering to mislead model predictions. In this paper, we propose a transformation-specific smoothing framework TPC, which provides tight and scalable robustness guarantees for point cloud models against semantic transformation attacks. We first categorize common 3D transformations into three categories: additive (e.g., shearing), composable (e.g., rotation), and indirectly composable (e.g., tapering), and we present generic robustness certification strategies for all categories respectively. We then specify unique certification protocols for a range of specific semantic transformations and their compositions. Extensive experiments on several common 3D transformations show that TPC significantly outperforms the state of the art. For example, our framework boosts the certified accuracy against twisting transformation along z-axis (within 20$^\circ$) from 20.3$\%$ to 83.8$\%$. Codes and models are available at https://github.com/chuwd19/Point-Cloud-Smoothing.
BibTeX
@inproceedings{chu2022tpc,
  author = {Wenda Chu and Linyi Li and Bo Li},
  title = {{TPC: Transformation-Specific Smoothing for Point Cloud Models}},
  booktitle = {International Conference on Machine Learning},
  year = {2022},
  url = {https://openreview.net/forum?id=18B9uqaRB3d},
}
Figure