Adversarial Examples For Fun And Profit

Try to generate adversarial images for fun and profit.

What is this

You can generate adversarial images in your browser by the power of tensorflow.js.

Demo: https://adv-examples-fun.netlify.com/

Attention! The above site is heavy. If you see a blank page, reload your browser.

Attacks

You can try the below attaks.

• FGSM
• DeepFool
• NewtonFool

Issues

Currently, the implementation of DeepFool attack is slightly modified from the original paper. This is because I found that the denominator of some formulas can be so small that the perturbation is too big. I am investigating this issue.

References

FGSM is described by

@misc{goodfellow2014explaining,
    title={Explaining and Harnessing Adversarial Examples},
    author={Ian J. Goodfellow and Jonathon Shlens and Christian Szegedy},
    year={2014},
    eprint={1412.6572},
    archivePrefix={arXiv},
    primaryClass={stat.ML}
}.

DeepFool is described by

@misc{moosavidezfooli2015deepfool,
    title={DeepFool: a simple and accurate method to fool deep neural networks},
    author={Seyed-Mohsen Moosavi-Dezfooli and Alhussein Fawzi and Pascal Frossard},
    year={2015},
    eprint={1511.04599},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}.

NewtonFool is described by

@inproceedings{10.1145/3134600.3134635,
author = {Jang, Uyeong and Wu, Xi and Jha, Somesh},
title = {Objective Metrics and Gradient Descent Algorithms for Adversarial Examples in Machine Learning},
year = {2017},
isbn = {9781450353458},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3134600.3134635},
doi = {10.1145/3134600.3134635},
booktitle = {Proceedings of the 33rd Annual Computer Security Applications Conference},
pages = {262–277},
numpages = {16},
keywords = {Adversarial Examples, Machine Learning},
location = {Orlando, FL, USA},
series = {ACSAC 2017}
}.

src/utils/data.ts, src/utils/model.ts and src/sagas/index.ts contain the modified version of the code from TensorFlow.js Example: Training MNIST. The license file is LICENSE.original.

Author

Catminusminus