TASAR: Transfer-based Attack on Skeletal Action Recognition
Authors: Yunfeng Diao, Baiqi Wu, Ruixuan Zhang, Ajian Liu, Xiaoshuai Hao, Xingxing Wei, Meng Wang, He Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | For exhaustive evaluation, we build the first large-scale robust S-HAR benchmark, comprising 7 S-HAR models, 10 attack methods, 3 S-HAR datasets and 2 defense models. Extensive results demonstrate the superiority of TASAR. Our benchmark enables easy comparisons for future studies, with the code available in the https://github.com/yunfengdiao/Skeleton-Robustness-Benchmark. |
| Researcher Affiliation | Academia | 1Hefei University of Technology 2 Institute of Automation Chinese Academy of Sciences 3 Beijing Academy of Artificial Intelligence 4 Beihang University 5 UCL Centre for Artificial Intelligence, Department of Computer Science, University College London |
| Pseudocode | Yes | Algorithm 1: Inference for Post-train Dual Bayesian Motion Attack |
| Open Source Code | Yes | Our benchmark enables easy comparisons for future studies, with the code available in the https://github.com/yunfengdiao/Skeleton-Robustness-Benchmark. [...] The source code and model checkpoint can also be found in the supplementary materials. |
| Open Datasets | Yes | Robust Bench HAR incorporates three popular S-HAR datasets: NTU 60 (Shahroudy et al., 2016) , NTU 120 (Liu et al., 2019) and HDM05(M uller et al., 2007). [...] all the datasets used in this paper are open dataset and are available to the public. |
| Dataset Splits | Yes | Due to variations in data pre-processing settings among S-HAR classifiers (such as data requiring subsampling(Zhang et al., 2019b)), we unify the data format following Wang et al. (2023). For NTU60 and NTU120, we subsample frames to 60. For HDM05, we segment the data into 60-frame samples(Diao et al., 2021). |
| Hardware Specification | Yes | The experimental platform used in this study is equipped with an AMD EPYC 7542 32-Core CPU operating at a clock speed of 2039.813 GHz, four NVIDIA Ge Force RTX 3090 GPUs, and 24 GB of memory per GPU. |
| Software Dependencies | No | The proposed method was implemented using the open-source machine learning framework Py Torch. No specific version number for PyTorch or Python is provided. |
| Experiment Setup | Yes | During the post-train, we set a learning rate of 0.03 with five epochs. We use SGAHMC optimizers (Springenberg et al., 2016), within itτ is automatically chosen, the friction coefficient F = 10 5 and Mθ = 30 steps for sampling. During the dual Bayesian optimization, we set γ = 0.1/ θ 2 and perform training for 5 epochs with a learning rate of 0.03. Additionally, we always set λε,σ = 0.001. When performing attacks, we set σ = 0.009 for models with SWAG. The w1,w2 and w3 are set as 0.8, 0.1 and 0.1. For a fair comparison, we run 200 iterations for all attacks under l norm-bounded perturbation of size 0.01. |