Quality-Preserving Imperceptible Adversarial Attack on Skeleton-based Human Action Recognition

Code for the paper "Quality-Preserving Imperceptible Adversarial Attack on Skeleton-based Human Action Recognition"

Environment Setup

Please create an appropriate environment using conda:

conda env create -f environment.yml

conda activate QPAttack

Download Data

100STYLE dataset can be downloaded from 100STYLE and HDM05 dataset can be downloaded from HDM05. Once downloaded, please extract it into ./dataset.

Please follow the frame-cut.csv to segment the long sequence of motions in the 100STYLE dataset. For HDM05 dataset, we follow the pre-processing precedure in SMART.

Methodology

Results

Online User Study (test if you can tell the adversarial motions from our method)