LogoStyleFool: Vitiating Video Recognition Systems via Logo Style Transfer
Authors: Yuxin Cao, Ziyu Zhao, Xi Xiao, Derui Wang, Minhui Xue, Jin Lu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results substantiate the overall superiority of Logo Style Fool over three state-ofthe-art patch-based attacks in terms of attack performance and semantic preservation. Meanwhile, Logo Style Fool still maintains its performance against two existing patch-based defense methods. Experiments Experimental Setup Datasets and Models. We choose UCF-101 (Soomro, Zamir, and Shah 2012) and HMDB-51 (Kuehne et al. 2011), two datasets that are popularly used in video adversarial attacks, to verify the attack performance. |
| Researcher Affiliation | Collaboration | Yuxin Cao1*, Ziyu Zhao2*, Xi Xiao1 , Derui Wang3, Minhui Xue3, Jin Lu4 1 Shenzhen International Graduate School, Tsinghua University, China 2 Fan Gongxiu Honors College, Beijing University of Technology, China 3 CSIRO s Data61, Australia 4 Ping An Technology (Shenzhen) Co., Ltd., China |
| Pseudocode | Yes | Algorithm 1: Logo Style Fool. |
| Open Source Code | Yes | The source code is available at https://github.com/ziyuzhao-zzy/Logo Style Fool. |
| Open Datasets | Yes | We choose UCF-101 (Soomro, Zamir, and Shah 2012) and HMDB-51 (Kuehne et al. 2011), two datasets that are popularly used in video adversarial attacks, to verify the attack performance. |
| Dataset Splits | No | The paper states: 'We beforehand trained the two models on two datasets.' and mentions 'test data split' but does not provide specific percentages or counts for training, validation, or test splits. It implicitly uses a test set for evaluation but the exact split ratios are not explicitly mentioned. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for experiments (e.g., GPU/CPU models, memory, or cloud instance specifications). |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We set the query limit as 3 105. The other parameters of benchmarks are set as their default values. ... λc, λs and λtv represent weight coefficients. ... µa and µd are the area and distance coefficients balancing the penalties on the logo size and the distance from the logo to the corner. ... η denotes the step size. ... ε stands for the perturbation threshold. |