Summary of Tasar: Transfer-based Attack on Skeletal Action Recognition, by Yunfeng Diao et al.
TASAR: Transfer-based Attack on Skeletal Action Recognition
by Yunfeng Diao, Baiqi Wu, Ruixuan Zhang, Ajian Liu, Xiaoshuai Hao, Xingxing Wei, Meng Wang, He Wang
First submitted to arxiv on: 4 Sep 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper investigates transfer-based attacks in Skeleton-based Human Activity Recognition (S-HAR), a crucial subfield of human-computer interaction. Recent work has shown that adversarial attacks can expose potential security concerns and provide valuable tools for model robustness testing. However, existing S-HAR attacks have weak adversarial transferability, and the reasons behind this phenomenon are largely unknown. The authors characterize the loss function to identify a prominent indicator of poor transferability: low smoothness. They propose a new attack method, TASAR (Transfer-based Attack on Skeletal Action Recognition), which smoothen’s the loss function during adversarial example computation. Unlike existing methods, TASAR incorporates motion dynamics into the Bayesian attack, disrupting spatial-temporal coherence in S-HARs. The authors build a large-scale robust S-HAR benchmark, comprising 7 models, 10 attack methods, 3 datasets, and 2 defense models. Experimental results demonstrate the superiority of TASAR. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research paper looks at how to make attacks on systems that recognize human actions (like recognizing you’re walking or running) more powerful. Currently, these attacks are not very good at transferring between different systems, and researchers don’t understand why. The authors identify a key factor in this weakness: the lack of smoothness in the way the attack works. They propose a new type of attack that incorporates this idea, which they call TASAR. Unlike other attacks, TASAR takes into account how people move over time, making it more effective. To test their approach, the authors created a large dataset of human activities and used it to compare different attacks. The results show that TASAR is much better than existing methods. |
Keywords
* Artificial intelligence * Activity recognition * Loss function * Transferability
