Dynamics-aware Adversarial Attack of 3D Sparse Convolution Network

An Tao, Yueqi Duan, He Wang, Ziyi Wu, Pengliang Ji, Haowen Sun, Jie Zhou, Jiwen Lu

TL;DR

This paper introduces a novel "Leaded Gradient Method" (LGM) to address a critical vulnerability in 3D sparse convolution networks: their dynamic architecture. LGM effectively "leads" adversarial attacks by accounting for how the network's structure changes during the attack, significantly improving attack performance compared to existing methods.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are susceptible to adversarial attacks, where small, imperceptible perturbations to input data can cause misclassifications. Most adversarial attack methods assume a fixed network architecture throughout the attack process. However, modern networks, particularly those designed for efficiency like 3D sparse convolution networks used in 3D scene analysis, often have dynamic architectures. These networks adapt their structure based on the input data to improve computational efficiency. This dynamic nature creates a "lagged gradient" problem: an attack designed for the network at one step becomes less effective in the next step because the network's architecture has changed. This paper tackles this challenge by developing an adversarial attack specifically designed for these dynamically evolving networks.

Methods

The core contribution is the Leaded Gradient Method (LGM). Traditional adversarial attacks use gradients calculated based on a fixed network. LGM, however, reformulates the gradient calculation to be "dynamics-aware."

Here's a breakdown of the key ideas:

  • Modeling Sparse Convolution: The authors first model 3D sparse convolution operations in a way that allows for gradient calculation even on voxels that might become empty or occupied due to perturbations. This involves introducing an "occupancy function" to represent voxel existence.
  • Soft Occupancy: To overcome the non-differentiable nature of hard occupancy (a voxel either exists or it doesn't), LGM introduces a differentiable soft occupancy function. This function estimates the likelihood of a voxel being occupied based on the proximity of input points. They propose a novel sigmoid-like function for this, which they show outperforms existing methods like bilinear interpolation and radial basis functions.
  • Leaded Gradient: The crucial innovation is the "leaded gradient." By incorporating the changes in occupancy values (which reflect the dynamic architecture) into the gradient calculation, LGM ensures that the learned perturbations are more effective for the next state of the network, not just the current one. This effectively "leads" the attack.
  • End-to-End Attack: The method is designed for end-to-end attack, meaning it can directly perturb input point cloud coordinates and propagate gradients through the entire network, including the voxelization and devoxelization steps, which are often non-differentiable.

Datasets

The paper evaluates its method on several standard 3D datasets for both semantic segmentation and classification tasks:

  • ScanNet: Large-scale indoor 3D dataset for semantic segmentation.
  • S3DIS: 3D indoor scans for semantic segmentation.
  • SemanticKITTI: Large-scale outdoor driving-scene 3D dataset for semantic segmentation.
  • ModelNet40: Dataset of 3D CAD models for object classification.

Key Insights and Ideas

  • Dynamic Architectures are a Major Vulnerability: The paper highlights that the assumption of fixed network architectures, common in adversarial attack research, is a critical blind spot when dealing with modern, efficient networks like 3D sparse convolutions.
  • "Lagged Gradient" is the Core Problem: The dynamic nature of these networks causes a temporal mismatch between the attack's gradient calculation and the network's actual state, rendering attacks less effective.
  • Soft Occupancy is Key to Differentiable Dynamics: Creating a differentiable approximation of voxel occupancy is essential for backpropagating gradients through the dynamic architectural changes.
  • "Leading" the Attack is More Effective: By anticipating and incorporating architectural shifts, the LGM attack is significantly more potent than its "dynamics-unaware" counterparts.

Key Results

The LGM attack demonstrates impressive performance across various datasets and tasks:

  • Semantic Segmentation: On ScanNet and S3DIS datasets, LGM achieves an average drop in mIoU (mean Intersection over Union) that is approximately 20% greater than dynamics-unaware methods like FGM. For example, on ScanNet with a budget of 0.05m, LGM achieves an mIoU of 3.83% compared to FGM's 8.70%.
  • SemanticKITTI: While the performance drop is not as pronounced as on ScanNet/S3DIS, LGM still outperforms FGM. This is attributed to SemanticKITTI's point features including coordinates, which aids gradient back-propagation.
  • Object Classification: On ModelNet40, LGM consistently outperforms FGM and random perturbation across different attack budgets. For instance, with a budget of 0.1m, LGM achieves 19.12% accuracy, while FGM achieves 25.85%.
  • Comparison to State-of-the-Art: LGM also outperforms recent point cloud perturbation attacks like PointPert and KNNPert, which do not account for dynamic network changes.
  • Qualitative Results: Visualizations show that even with imperceptible perturbations (e.g., 0.5-1 cm), LGM can cause significant misclassifications, highlighting a substantial security risk.

Limitations and Future Work

  • Transferability to Other Dynamic Architectures: The current LGM is specifically tailored for 3D sparse convolution networks. The authors suggest future work could explore adapting the "leaded gradient" concept to other types of networks with dynamic architectures (e.g., conditional computation, Dynamic DNNs).
  • Leveraging More Advanced Attack Methods: LGM is built upon the Fast Gradient Method (FGM). Future research could investigate how to integrate more sophisticated attack strategies (beyond FGM) with the leaded gradient approach.
  • Negative Societal Impact: The paper acknowledges that the ability to generate highly effective, imperceptible adversarial examples raises concerns about negative societal impacts, particularly in security applications. Further research is needed to explore defenses against such attacks.

Summary updated

Abstract

In this paper, we investigate the dynamics-aware adversarial attack problem in deep neural networks. Most existing adversarial attack algorithms are designed under a basic assumption -- the network architecture is fixed throughout the attack process. However, this assumption does not hold for many recently proposed networks, e.g. 3D sparse convolution network, which contains input-dependent execution to improve computational efficiency. It results in a serious issue of lagged gradient, making the learned attack at the current step ineffective due to the architecture changes afterward. To address this issue, we propose a Leaded Gradient Method (LGM) and show the significant effects of the lagged gradient. More specifically, we re-formulate the gradients to be aware of the potential dynamic changes of network architectures, so that the learned attack better "leads" the next step than the dynamics-unaware methods when network architecture changes dynamically. Extensive experiments on various datasets show that our LGM achieves impressive performance on semantic segmentation and classification. Compared with the dynamic-unaware methods, LGM achieves about 20% lower mIoU averagely on the ScanNet and S3DIS datasets. LGM also outperforms the recent point cloud attacks.
BibTeX
@misc{tao2021dynamics,
  author = {An Tao and Yueqi Duan and He Wang and Ziyi Wu and Pengliang Ji and Haowen Sun and Jie Zhou and Jiwen Lu},
  title = {{Dynamics-aware Adversarial Attack of 3D Sparse Convolution Network}},
  year = {2021},
  eprint = {2112.09428},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2112.09428},
}
Figure