Towards Transferable Adversarial Attacks with Centralized Perturbation

Authors: Shangbo Wu, Yu-an Tan, Yajie Wang, Ruinan Ma, Wencong Ma, Yuanzhang Li

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments demonstrate that by dynamically centralizing perturbation on dominating frequency coefficients, crafted adversarial examples exhibit stronger transferability, and allowing them to bypass various defenses.
Researcher Affiliation Academia 1School of Cyberspace Science and Technology, Beijing Institute of Technology 2School of Computer Science and Technology, Beijing Institute of Technology {shangbo.wu, tan2008, wangyajie19, ruinan, wencong.ma, popular}@bit.edu.cn
Pseudocode Yes Algorithm 1: Centralized Adversarial Perturbation Input: Original image x, ground truth label y, source model F with loss J . Parameters: Iteration T, size of perturbation ε, learning rate β, quantization ratios r Y , r Cb, r Cr and corresponding quantization matrices QY , QCb, QCr (denoted as Qs for brevity) each Y/Cb/Cr channel. Output: xadv T . 1: Let step size α ε/T, Q0 = 1. 2: for t = 0 to T 1 do 3: Acquire gradient from F as x J (xadv t , y). 4: Perturbation δt = α sign( x J (xadv t , y)). 5: Optimize δt by Equation 5 and 7 to acquire δ t, and clip with respect to ε. 6: Acquire intermediate xadv t = x + δ t. 7: Update Qts by passing the quantized xadv t (through the same Equation 5 and 7) to F and solving the optimization via Equation 6. 8: end for 9: return xadv T = x + δ T .
Open Source Code No The paper references third-party code for pre-trained models ('PyTorch Image Models. https://github.com/huggingface/pytorch-image-models'), but does not explicitly state that the source code for their own methodology is available or provide a link to it.
Open Datasets Yes The dataset from the Neur IPS 2017 Adversarial Learning Challenge (Kurakin et al. 2018) is used, consisting of 1000 images from Image Net with shape (3, 224, 224).
Dataset Splits No The paper mentions the dataset used but does not provide specific training, validation, or test splits for it within the context of their experiments.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions 'PyTorch Image Models' as a source for pre-trained models but does not provide specific version numbers for it or any other ancillary software dependencies.
Experiment Setup Yes Hyper-parameters. Attacks are ℓ -bounded. ε = 8/255. They run for iteration T = 10, 20, 50. Learning rate of the Adam optimizer β = 0.1. Quantization ratios are set to r Y = 0.9, r Cb = 0.05, r Cr = 0.05, with a total quantization rate of 1/3.