Skip to content

Latest commit

 

History

History
46 lines (29 loc) · 2.66 KB

README.md

File metadata and controls

46 lines (29 loc) · 2.66 KB

NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks

Data and model can be found here: data and model.

Please download the data&&model and unzip them to './cifar-data' and './all_models'

Below is Table 1 from our paper, where we show the robustness of each accepted defense to the adversarial examples we can construct:

Defense Dataset Distance Success rate
ADV-TRAIN Madry et al. (2018) CIFAR 0.031 (linf) 47.9%
ADV-BNN Liu et al. (2019) CIFAR 0.035 (linf) 75.3%
THERM-ADV Buckman et al. (2018)Madry et al. (2018) CIFAR 0.031 (linf) 91.2%
CAS-ADV Na et al. (2018) CIFAR 0.031 (linf) 97.7%
ADV-GAN Wang & Yu (2019) CIFAR 0.015 (linf) 98.3%
LID Ma et al. (2018) CIFAR 0.031 (linf) 100.0%
THERM Buckman et al. (2018) CIFAR 0.031 (linf) 100.0%
SAP Dhillon et al. (2018) CIFAR 0.031 (linf) 100.0%
RSE Liu et al. (2018) CIFAR 0.031 (linf) 100.0%
GUIDED DENOISER (Liao et al., 2018) ImageNet 0.031 (linf) 95.5%
RANDOMIZATION Xie et al. (2018) ImageNet 0.031 (linf) 96.5%
INPUT-TRANS Guo et al. (2018) ImageNet 0.005 (l2) 100.0%
PIXEL DEFLECTION Prakash et al. (2018) ImageNet 0.031 (linf) 100.0%

Paper

Paper available in Arxiv. If you feel it helpful, please cite our work.

@inproceedings{ Li2019NATTACKLT, title={NATTACK: Learning the Distributions of Adversarial Examples for an Improved Black-Box Attack on Deep Neural Networks}, author={Yandong Li and Lijun Li and Liqiang Wang and Tong Zhang and Boqing Gong}, year={2019} }

Source code

This repository contains our implemenation of the black-box attack algorithm described in our paper, six defense methods (SAP, LID, RANDOMIZATION, INPUT-TRANS, THERM, and THERM-DAV) borrowed from the code of Anish et al. (2018), two defended models (GUIDED DENOISER and PIXEL DEFLECTION) based on the code of Athalye & Carlini, (2018), and two defended models (RSE and CAS-ADV) from the original papers.