Understanding and Defending Patched-based Adversarial Attacks for Vision Transformer

Authors: Liang Liu, Yanan Guo, Youtao Zhang, Jun Yang

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental First, we experimentally observe that adversarial patches only activate in a few layers and become lazy during attention updating. According to experiments, we study how a small adversarial patch perturbates the whole model. In this work, we first design two experiments to deeply understand why such a small-size patch can crash the entire Vi T model.
Researcher Affiliation Academia 1Department of ECE, University of Pittsburgh 2Department of CS, University of Pittsburghy.
Pseudocode Yes Algorithm 1 ARMOR
Open Source Code No The paper does not provide an explicit statement or link to the open-source code for their proposed method (ARMOR).
Open Datasets Yes We choose 384 384 3 images randomly sampled from Image Net 2012 (Deng et al., 2009).
Dataset Splits No The paper mentions using "100 images to learn the detection threshold" but does not provide specific train/validation/test dataset splits, percentages, or explicit predefined splits.
Hardware Specification Yes Our code runs on a 4090 GPU, while Certified-Patch and Smooth-Vi T use a V100 GPU.
Software Dependencies No All codes are written in Python and Py Torch (Paszke et al., 2019) Platform.
Experiment Setup Yes In the experiment, we use 100 images to learn the detection threshold, τ, for each model. We set the number of detection as Nd = 5. We set the perturbation radius of the adversarial patch to 0.8, which means the value of each pixel can be altered to 204/255. The size of adversarial patches is equal to the patch size of the model, e.g., it is 32 32 pixels for Vi T/32 and 16 16 pixels for Vi T/16.