SVAttack: Spatial-Viewpoint Transfer Attack on Graph Convolutional Skeleton Action Recognition

Zhize Wu

doi:10.1109/tifs.2026.3734764

SVAttack: Spatial-Viewpoint Transfer Attack on Graph Convolutional Skeleton Action Recognition

2026-09-17

0

PRE

AI

DOI：10.1109/tifs.2026.3734764

Original

Original request for help

Save

Abstract

En 中文

Skeleton-based action recognition with graph convolutional networks has achieved remarkable success and is increasingly deployed in safety-critical applications. This necessitates research on possible attack methods and their limitations. The existing adversarial attacks on skeleton-based models suffer from limited transferability across architectures and often exhibit noticeable perceptual distortion. In this work, we investigate the architectural inductive biases that fundamentally constrain transferability in skeleton-based action recognition. Based on a structural and gradient-level analysis, we propose (1) a spatial gradient damper to suppress model-specific spatial biases and (2) a viewpoint-based constraint mechanism to improve perceptual imperceptibility. Extensive experimental results on 3 benchmark datasets, 9 skeleton-based action recognition models, 7 attack methods, and 2 defense methods demonstrate that the proposed approach significantly enhances transferability while maintaining visually plausible skeleton motions. It establishes a new state of the art for transferable adversarial attacks on skeleton-based action recognition models. Our code is available at https://github.com/deep-wu/SVAttack.

Keywords:

Architectural Inductive Biases

Graph Convolution

Transfer Attack

Action Recognition

SVAttack: Spatial-Viewpoint Transfer Attack on Graph Convolutional Skeleton Action Recognition

Abstract

Journal

Organization

Cited Papers