TL;DR
This paper introduces a novel black-box adversarial attack method for 3D mesh classification networks. By training an "imitating network" that learns to predict the outputs of a target network using random walks on the mesh, they can generate imperceptible mesh modifications that cause misclassification, demonstrating the vulnerability of current SOTA models.
What Problem Does This Paper Address?
While adversarial attacks on 2D images are well-studied, the domain of 3D mesh classification has seen less exploration. 3D meshes, a common representation for surfaces in computer graphics, present unique challenges for adversarial attacks. Unlike point clouds, modifying mesh vertices can easily lead to noticeable geometric distortions or topological changes, making it difficult to create attacks that are both effective and visually imperceptible. This paper aims to fill this gap by proposing a general, black-box adversarial attack specifically for mesh classification networks.
Methods
The core of the proposed method is an "imitating network" trained to mimic the behavior of a target mesh classification network. This imitating network takes a mesh and a random walk on its surface as input and outputs a prediction vector. The random walk is crucial as it helps gather geometric information from the mesh, and the imitating network learns to associate these geometric features with the target network's predictions.
The attack process is black-box, meaning it only has access to the target network's predictions, not its internal architecture or gradients. When attacking a specific network, the imitating network is trained on pairs of meshes and their corresponding prediction vectors from the target network. The loss function used is Kullback-Leibler Divergence (KLD) to match the full probability distributions of predictions, rather than just the predicted class.
Once trained, the imitating network is used to generate adversarial meshes. The process involves extracting random walks from the mesh and using the imitating network to predict the target network's output. Gradients are then computed based on the KLD loss between the imitating network's prediction and the original class. The mesh vertices along the walk are updated in the opposite direction of these gradients. This approach ensures that vertices with higher influence on the target network's classification are modified more, spreading the changes across the mesh to maintain visual imperceptibility.
Datasets
The paper evaluates its method on two widely used 3D mesh datasets:
- SHREC11: Contains 600 meshes across 30 classes.
- ModelNet40: A larger dataset with 12,311 CAD meshes across 40 categories.
Key Insights and Ideas
- Imitating Network for Black-Box Attacks: The concept of training an "imitating network" to learn the prediction function of a target network is a novel approach for black-box adversarial attacks on meshes.
- Random Walks as Geometric Probes: Random walks are shown to be an effective tool for exploring mesh surfaces and extracting relevant geometric information that can be used to understand and attack classification networks.
- Gradient-Based Modification for Imperceptibility: By using gradients derived from the imitating network's loss and updating vertices in the opposite direction, the attack can prioritize modifications to influential regions while still spreading changes across the mesh, achieving both misclassification and visual similarity.
- Architecture-Specific Insights: The attack's ability to highlight different regions of the mesh for different networks provides valuable insights into how these networks make their classifications.
Key Results
- Significant Accuracy Drop: The proposed attack successfully reduces the accuracy of several state-of-the-art mesh classification networks on both SHREC11 and ModelNet40 datasets. For example, PD-MeshNet's accuracy dropped from 99.7% to 18.3% on SHREC11.
- Visual Imperceptibility: Qualitative evaluations and a user study confirm that the modified meshes are visually similar to the original ones. In a user study, 69% of participants believed the attacked meshes belonged to the same class as the original, while only 4% agreed with the network's misclassification.
- Comparison to White-Box Attacks: The black-box attack achieves comparable or better results than existing white-box attacks, demonstrating its effectiveness. For instance, it achieved a lower post-attack accuracy for MeshNet on ModelNet40 compared to a white-box method.
- Amount of Modification: The L2 distortion analysis shows that the amount of modification required is relatively consistent across different networks and classes, though some classes (e.g., man-made objects) are harder to modify imperceptibly.
Limitations and Future Work
- Noisy Modifications for Man-Made Objects: The paper acknowledges that for certain classes, particularly man-made objects with sharp geometric features, the modifications can be perceived as "noisy" by humans, even if they fool the network. This is attributed to the L2 distance metric not capturing geometric property violations.
- Future Work - Targeted Attacks: The authors suggest extending the attack to targeted attacks, where the goal is to make the network classify the attacked model as a specific target class, rather than just any incorrect class. Preliminary studies in this direction have shown promising results.
- Future Work - Improving Robustness: The findings of this paper can inform the development of more robust mesh classification networks by highlighting their vulnerabilities.
Summary updated