Adversarial Bone Length Attack on Action Recognition

Authors: Nariki Tanaka, Hiroshi Kera, Kazuhiko Kawamoto2335-2343

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted experiments on the NTU RGB+D and HDM05 datasets and demonstrate that the proposed attack successfully deceived models with sometimes greater than 90% success rate by small perturbations.
Researcher Affiliation Academia Nariki Tanaka,1 Hiroshi Kera,2 Kazuhiko Kawamoto,2 1 Graduate School of Science and Engineering, Chiba University 2 Graduate School of Engineering, Chiba University {ntanaka, kera}@chiba-u.jp, kawa@faculty.chiba-u.jp
Pseudocode Yes Algorithm 1: Pseudocode of adversarial bone length attack
Open Source Code No The paper provides links to the official code for the *target models* (ST-GCN and SGN) but not to the authors' own implementation of the proposed adversarial bone length attack methodology.
Open Datasets Yes We used the NTU RGB+D (Shahroudy et al. 2016) and HDM05 (M uller et al. 2007) datasets, which are 3D skeleton action datasets.
Dataset Splits Yes We randomly divided samples of each class into a training set (80%), validation set (10%), and testing set (10%).
Hardware Specification Yes All experiments were conducted using an Intel Core i7-6850K CPU and TITAN RTX GPU.
Software Dependencies No The paper mentions using Python and related libraries for deep learning (e.g., implied by ST-GCN, SGN, PyTorch/TensorFlow frameworks), but does not specify exact version numbers for any software dependencies.
Experiment Setup Yes The maximum number of iterations of the PGD was set to 50. The step size was set to α = 0.01, as in (Liu, Akhtar, and Mian 2020).