Boosting 3D Adversarial Attacks with Attacking On Frequency

Binbin Liu, Jinlai Zhang, Lyujie Chen, Jihong ZhuIEEE Access 2022 · OpenReview

TL;DR

This paper introduces AOF, a novel adversarial attack for 3D point clouds that targets the low-frequency components (LFC) of the data. By focusing on these intrinsic shape features, AOF significantly improves the transferability of adversarial attacks to unseen models and makes them more robust against common defense mechanisms.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are increasingly used for 3D vision tasks, but they are susceptible to adversarial attacks. Existing methods for generating adversarial 3D point clouds often suffer from poor transferability (i.e., attacks crafted for one model don't work well on another) and are easily detected by defense systems. This is partly because these attacks indiscriminately perturb all parts of the point cloud, including "noisy" features that are model-specific, rather than focusing on the core, intrinsic properties of the object's shape. The paper aims to bridge this gap by developing an attack that leverages the inherent structural information within point clouds.

Methods

The core idea behind AOF is to exploit the low-frequency components (LFC) of a point cloud, which the authors hypothesize represent the fundamental shape of an object. They achieve this by:

  1. Graph Fourier Transform: Applying a graph Fourier transform to the point cloud to decompose it into LFC and high-frequency components (HFC). This is necessary because standard Fourier transforms, used in image domains, don't directly apply to irregular point cloud structures.
  2. LFC-Focused Perturbation: Crafting adversarial samples by specifically optimizing perturbations on the LFC of the point cloud. This is done by introducing an additional adversarial loss term that targets the LFC, in addition to the standard adversarial loss on the entire point cloud.
  3. Combined Loss: The optimization process uses a weighted combination of the adversarial loss on the full point cloud and the adversarial loss on its LFC. This allows the attack to be effective while prioritizing changes to the object's fundamental shape.

Datasets

The experiments were conducted using the ModelNet40 dataset, a widely used benchmark for 3D object classification. This dataset was used for training the victim networks and evaluating the proposed attack.

Key Insights and Ideas

  • Low-Frequency Components as Intrinsic Features: The paper posits that the LFC of a point cloud captures the essential, stable shape information of an object, analogous to how low-frequency components define the overall structure in images.
  • Transferability Through Intrinsic Features: By focusing attacks on these intrinsic LFCs, AOF can generate adversarial examples that are more likely to fool different models, as these core features are less model-specific than high-frequency perturbations.
  • Deformation vs. Outliers: AOF-generated adversarial point clouds tend to cause subtle deformations rather than introducing outliers, making them harder for defense mechanisms to detect and remove.
  • Frequency Domain Analysis: The paper highlights the potential of frequency domain analysis for understanding and improving adversarial attacks in 3D vision, a direction less explored than in image domains.

Key Results

  • Improved Transferability: AOF significantly outperforms state-of-the-art (SOTA) attacks like 3D-Adv, kNN, and AdvPC in terms of transferability across different 3D neural networks. For example, in Table 2, AOF shows transferability rates up to ~70% higher than baselines.
  • Robustness to Defenses: AOF demonstrates superior robustness against various SOTA 3D defense algorithms (SRS, SOR, DUP-Net, IF-Defense) compared to existing attacks. Table 3 shows AOF maintaining higher attack success rates even after defenses are applied.
  • Subtle Perturbations: Visualizations (Figure 6) show that AOF-generated adversarial samples exhibit more subtle deformations compared to the outliers often introduced by other methods.
  • Hyperparameter Sensitivity: Ablation studies showed that the choice of hyperparameter m (number of LFC eigenvectors) is crucial, with an optimal range balancing transferability and attack success. The hyperparameter γ (weighting of LFC loss) also plays a role in balancing white-box success rate and transferability.

Limitations and Future Work

  • Targeted Attacks: While AOF shows comparable targeted transferability to kNN, improving targeted attack transferability remains a challenging but promising area.
  • Imperceptibility: The authors suggest exploring imperceptibility by restricting perturbations to the HFC in future work, which could lead to even more stealthy attacks.
  • Spectral Analysis: The paper includes a spectral analysis of perturbations, noting differences in how attacks transfer to different models (e.g., PointNet vs. DGCNN), suggesting further avenues for research into the underlying reasons for transferability variations.

Summary updated

Abstract

Deep neural networks (DNNs) have been shown to be vulnerable to adversarial attacks. Recently, 3D adversarial attacks, especially adversarial attacks on point clouds, have elicited mounting interest. However, adversarial point clouds obtained by previous methods show weak transferability and are easy to defend. To address these problems, in this paper we propose a novel point cloud attack (dubbed AOF) that pays more attention on the low-frequency component of point clouds. We combine the losses from point cloud and its low-frequency component to craft adversarial samples. Extensive experiments validate that AOF can improve the transferability significantly compared to state-of-the-art (SOTA) attacks, and is more robust to SOTA 3D defense methods. Otherwise, compared to clean point clouds, adversarial point clouds obtained by AOF contain more deformation than outlier.
BibTeX
@inproceedings{liu2022boosting,
  author = {Binbin Liu and Jinlai Zhang and Lyujie Chen and Jihong Zhu},
  title = {{Boosting 3D Adversarial Attacks with Attacking On Frequency}},
  booktitle = {IEEE Access},
  year = {2022},
  url = {https://openreview.net/forum?id=GvzkiPTxe2},
}
Figure