Improving Integrated Gradient-based Transferable Adversarial Examples by Refining the Integration Path
Authors: Yuchen Ren, Zhengyu Zhao, Chenhao Lin, Bo Yang, Lu Zhou, Zhe Liu, Chao Shen
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments validate that Mu Mo DIG outperforms the latest IG-based attack by up to 37.3% and other state-of-the-art attacks by 8.4%. |
| Researcher Affiliation | Academia | 1Xi an Jiaotong University 2Information Engineering University 3Nanjing University of Aeronautics and Astronautics 4Zhejiang Lab EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology in prose and mathematical formulations within the 'Methodology' section, without structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Appx. & Code https://github.com/RYC-98/Mu Mo DIG |
| Open Datasets | Yes | Following many previous works (Wang and He 2021; Zhu et al. 2023; Long et al. 2022), 1k images from the ILSVRC2012 (Russakovsky et al. 2015) validation set are adopted in our experiments. |
| Dataset Splits | Yes | Following many previous works (Wang and He 2021; Zhu et al. 2023; Long et al. 2022), 1k images from the ILSVRC2012 (Russakovsky et al. 2015) validation set are adopted in our experiments. |
| Hardware Specification | Yes | All experiments are conducted on an RTX 4060 GPU with 8GB of VRAM. |
| Software Dependencies | No | The paper does not explicitly state specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | Following the common practice, for all attacks, we set the maximum attack iterations as K = 10, the maximum perturbation bound ̖ = 16, the step size ̖ = 1.6, the decay factor µ = 1.0 in the momentum. We set the position factor ̖ = 0.65 and the region number NR = 2 in LBQ. For a fair comparison, we set the number of total auxiliary inputs N = 6 at each iteration for all attacks. Specifically, for our Mu Mo DIG, we set NT = 6, NB = 1, and NI = 1 such that N = NT NB NI = 6. |