Enhancing Adversarial Transferability with Adversarial Weight Tuning

Authors: Jiahao Chen, Zhou Feng, Rui Zeng, Yuwen Pu, Chunyi Zhou, Yi Jiang, Yuyou Gan, Jinbao Li, Shouling Ji

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on a variety of models with different architectures on Image Net demonstrate that AWT yields superior performance over other attacks, with an average increase of nearly 5% and 10% attack success rates on CNN-based and Transformer-based models, respectively, compared to state-of-the-art attacks.
Researcher Affiliation Academia 1 College of Computer Science and Technology, Zhejiang University 2 Shandong Artificial Intelligence Institute 3 School of Mathematics and Statistics, Qilu University of Technology
Pseudocode Yes Algorithm 1: Adversarial Weight Tuning (AWT) attack
Open Source Code No The paper does not provide an explicit statement about releasing source code or a link to a repository for the methodology described.
Open Datasets Yes Evaluation is conducted on the Image Net-compatible dataset, which is widely utilized in prior work (Qin et al. 2022; Ge et al. 2023; Qiu et al. 2024).
Dataset Splits No The paper mentions using an "Image Net-compatible dataset... comprises 1,000 images", but it does not specify how these 1,000 images were split into training, validation, or test sets for the experiments.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific software details with version numbers (e.g., library names with versions like Python 3.8, PyTorch 1.9) needed to replicate the experiment.
Experiment Setup Yes Following the previous work (Ge et al. 2023; Qin et al. 2022; Qiu et al. 2024), we set the maximum perturbation ϵ = 16.0/255, the number of iterations T = 10, and the step size α = 1.6. For MI and NI, the decay factor µ = 1.0. For VMI, we set the number of sampled examples N = 20 and the upper bound of the neighborhood size β = 1.5 ϵ. For EMI, we set N = 11, the sampling interval bound η = 7, and use linear sampling. For the RAP attack, we set α = 2.0/255, K = 400, the inner iteration number T = 10, the late-start KLS = 100, and the size of neighborhoods ϵn = 16.0/255. For PGN, NCS and AWT, we set N = 20, the balanced coefficient δ = 0.5, and the upper bound ζ = 3.0 ϵ. For AWT alone, we set β = 0.005 and lr = 0.002 for surrogate models used for evaluation.