Can Adversarial Examples Be Parsed to Reveal Victim Model Information?

[Paper][Code][MSU OPTML Group]

Numerous adversarial attack methods have been developed to generate imperceptible image perturbations that can cause erroneous predictions of state-of-the-art machine learning (ML) models, in particular, deep neural networks (DNNs). Despite intense research on adversarial attacks, little effort was made to uncover 'arcana' carried in adversarial attacks. In this work, we ask whether it is possible to infer data-agnostic victim model (VM) information (i.e., characteristics of the ML model or DNN used to generate adversarial attacks) from data-specific adversarial instances. We call this 'model parsing of adversarial attacks' - a task to uncover 'arcana' in terms of the concealed VM information in attacks. We approach model parsing via supervised learning, which correctly assigns classes of VM's model attributes (in terms of architecture type, kernel size, activation function, and weight sparsity) to an attack instance generated from this VM. We collect a dataset of adversarial attacks across 7 attack types generated from 135 victim models (configured by 5 architecture types, 3 kernel size setups, 3 activation function types, and 3 weight sparsity ratios). We show that a simple, supervised model parsing network (MPN) is able to infer VM attributes from unseen adversarial attacks if their attack settings are consistent with the training setting (i.e., in-distribution generalization assessment). We also provide extensive experiments to justify the feasibility of VM parsing from adversarial attacks, and the influence of training and evaluation factors in the parsing performance (e.g., generalization challenge raised in out-of-distribution evaluation). We further demonstrate how the proposed MPN can be used to uncover the source VM attributes from transfer attacks, and shed light on a potential connection between model parsing and attack transferability.

Preparation

Change the path related arguments in path.py

TinyImageNet Preparation

cd TinyImageNet
bash run.sh

Pretrained DnCNN download

The pretrained DnCNN model is in GoogleDrive, same as the denoising-only objective trained model mentioned in "REVERSE ENGINEERING OF IMPERCEPTIBLE ADVERSARIAL IMAGE PERTURBATIONS".

Pipeline

Train victim model

python run.py --stage 0 --thread 3 --gpus 0,1,2,3,4,5,6,7

Get attack images

python run.py --stage 1 --thread 3 --gpus 0,1,2,3,4,5,6,7

Aggregate datas

python grep_data.py # for each attack strength
python grep_class_data.py # for mixed datasets

Train parsing model

python run.py --stage 2 --thread 3 --gpus 0,1,2,3,4,5,6,7

Evaluate parsing model

python run.py --stage 3 --thread 20 --gpus 0,1,2,3,4,5,6,7

Train and Evaluate parsing model with denoiser

python run.py --stage 2 --thread 3 --gpus 0,1,2,3,4,5,6,7 --denoise
python run.py --stage 3 --thread 5 --gpus 0,1,2,3,4,5,6,7 --denoise

Optional arguments in run.py

--stage STAGE    To decide which part of commands to execute. (0, 1, 2, 3)
--gpus GPUS      Run on which gpus. e.g.: --gpus 0,1,2,3
--thread THREAD  Number of commands running parallel in one gpu.
--denoise        Using denoiser when training attribute models. 
--debug          Generate commands without executing if tagged.

Citation

@article{yao2023can,
  title={Can Adversarial Examples Be Parsed to Reveal Victim Model Information?},
  author={Yao, Yuguang and Liu, Jiancheng and Gong, Yifan and Liu, Xiaoming and Wang, Yanzhi and Lin, Xue and Liu, Sijia},
  journal={arXiv preprint arXiv:2303.07474},
  year={2023}
}