Disclaimer: These are my personal notes on this paper. I am in no way related to this paper. All credits go towards the authors.

ADVBOX: A TOOLBOX TO GENERATE ADVERSARIAL EXAMPLES THAT FOOL NEURAL NETWORKS

Aug. 26, 2020 - Paper Link - Tags: Adversarial, Framework, Perturbation

Summary

Advbox is a toolbox used to generate adversarial examples that fool neural networks. It can also be used as a machine learning benchmarking tool.

The following frameworks are supported: Tensorflow, Keras, PyTorch, PaddlePaddle, Caffe2, and MxNet. A number of attack and defense methods are also supported.

Code

Notes

Focuses on perturbation attacks
Supported adversarial attacks (section 4.3):

FGSM
BIM
DeepFool
JSMA
CW
PGD

Supported defense algorithms (section 4.4.):

Feature Squeezing
Spatial Smoothing
Label Smoothing
Gaussian Augmentation
Adversarial Training
Thermometer Encoding

Uses Perceptron to evaluate tests. Supported attack and evaluation approaches:

CW
Gaussian Noise
Uniform Noise
Pepper Noise
Gaussian Blurs
Brightness
Rotations
Bad Weather

Interesting References

Analysis

Ideas

Citation: Goodman, Dou, et al. "Advbox: a toolbox to generate adversarial examples that fool neural networks." arXiv preprint arXiv:2001.05574 (2020).