Invisible Backdoor Attack against 3D Point Cloud Classifier in Graph Spectral Domain

Linkun Fan, Fazhi He, Tongzhen Si, Wei Tang, Bing LiAAAI 2024 · DOI

TL;DR

This paper introduces IBAPC, a novel method for creating invisible backdoor attacks on 3D point cloud classifiers. By manipulating the graph spectral domain, IBAPC implants malicious behavior that is imperceptible to human inspection while maintaining high attack efficiency, posing a significant threat to security-critical applications like autonomous driving.

What Problem Does This Paper Address?

3D point clouds are increasingly vital for applications demanding high accuracy and security, such as self-driving cars and facial recognition systems. However, these systems, often powered by Deep Neural Networks (DNNs), are vulnerable to sophisticated attacks. While adversarial attacks aim to fool models during inference, backdoor attacks are more insidious: they are implanted during the training phase and lie dormant until a specific "trigger" is present in the input. Previous 3D backdoor attacks, while effective, often introduce noticeable deformations, making them easily detectable by human observers. This paper tackles the challenge of creating invisible backdoor attacks for 3D point clouds, where the implanted trigger is undetectable both visually and through standard analysis.

Methods

The core innovation of IBAPC lies in its approach to implanting the backdoor trigger in the graph spectral domain rather than the spatial domain. Here's a breakdown of their method:

  • Graph Spectral Domain Manipulation: The authors leverage the Graph Fourier Transform (GFT) to convert 3D point clouds into their spectral representations. They hypothesize that manipulating these spectral signals can induce both global shape variations and local point perturbations, making the resulting deformation harder to pinpoint.
  • Backdoor Implanting Function (Tε(·)): A novel function, Tε(·), is proposed. This function takes a clean point cloud, transforms it to the graph spectral domain, adds a learned backdoor trigger (represented by a spectral perturbation vector ξ), and then transforms it back to the spatial domain to produce the backdoored point cloud.
  • Alternating Training Procedure: To learn both the optimal backdoor trigger (ξ) and the victim 3D DNN's parameters (θ) simultaneously, IBAPC employs an alternating training strategy. The parameters of the implanting function and the DNN are updated iteratively, ensuring the DNN learns to misclassify triggered inputs while maintaining accuracy on clean data and minimizing the visual difference between clean and backdoored samples.
  • Stealthiness Objective: A key objective is to minimize the Euclidean distance between the original point cloud and the backdoored point cloud (D(Pi, Tε(Pi))). This explicit regularization term in the loss function drives the deformation to be as subtle as possible.

Datasets

The paper evaluates IBAPC using standard 3D point cloud datasets:

  • ModelNet40 (MN40): A widely used dataset for 3D object classification, containing 40 object categories.
  • ModelNet10 (MN10): A subset of ModelNet40 with 10 object categories.
  • ShapeNetPart (SNP): A dataset for 3D part segmentation, used here for classification evaluation.

The authors uniformly sample 1,024 points from each object and rescale them to a unit cube for consistency.

Key Insights and Ideas

  • Spectral Domain is Key to Stealth: The paper strongly argues that manipulating the graph spectral domain is crucial for achieving invisibility. By perturbing frequencies, the attack can distribute the deformation across the entire point cloud in a way that is less conspicuous than localized spatial attacks.
  • Global and Local Deformation Synergy: The spectral approach allows the attack to influence both the overall shape (low frequencies) and fine-grained details (high frequencies) of the point cloud. This combined effect makes the resulting deformation more natural and harder to detect.
  • Learned Triggers Outperform Manual Design: Instead of manually designing triggers, learning them through an alternating optimization process allows for more effective and stealthy perturbations that are tailored to the specific victim DNN.
  • Stealthiness vs. Attack Efficiency Trade-off: While IBAPC might not always achieve the absolute highest Attack Success Rate (ASR) compared to some less stealthy methods, the paper demonstrates that the slight drop in ASR is a worthwhile trade-off for significantly improved stealthiness.

Key Results

  • SOTA Stealthiness: IBAPC achieves state-of-the-art stealthiness across three metrics: objective distance measurement (L2 distance), subjective human evaluation, and graph spectral signal residual. For instance, the L2 distance of IBAPC was found to be only 14.29% of the second-best algorithm on ModelNet40 with PointNet++.
  • Competitive Attack Efficiency: IBAPC achieves competitive Attack Success Rates (ASR), often only a few percentage points lower than the best-performing attacks, while offering vastly superior stealth.
  • Robustness to Defenses: IBAPC demonstrates resistance to several common defense mechanisms, including STRIP, saliency-based defenses, and data augmentation, highlighting its effectiveness against current detection methods.
  • Human Inspection Success: In human evaluation studies, participants identified IBAPC's backdoored point clouds as most similar to the benign ones in 66.79% of cases, significantly outperforming other attacks.

Limitations and Future Work

  • Clean Label Backdoors Not Addressed: The authors acknowledge that their current work focuses on attacks where the backdoored sample is still associated with its original label. The "invisible 3D backdoor attack under clean label" remains an open challenge for future research.
  • Potential for Further Optimization: While IBAPC is highly stealthy, the paper mentions that the "altered label will alarm defender." This suggests that future work could explore methods to further mask the misclassification itself, perhaps by ensuring the misclassified output is still plausible or falls within a similar category.
  • Exploration of Spectral Domain: The paper's success in the spectral domain opens avenues for exploring other spectral manipulation techniques or different graph representations for even more sophisticated attacks or defenses.

Summary updated

BibTeX
@inproceedings{fan2024invisible,
  author = {Linkun Fan and Fazhi He and Tongzhen Si and Wei Tang and Bing Li},
  title = {{Invisible Backdoor Attack against 3D Point Cloud Classifier in Graph Spectral Domain}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i19.30099},
}
Figure