search

erko / nips17-defense

Submission of NIPS 2017: Defense Against Adversarial Attack by Yerkebulan Berdibekov Edit Add topics
Apache License 2.0
1 stars 0 forks source link

Resuts/Contribution PDF #2

Closed pGit1 closed 6 years ago

pGit1 commented 6 years ago

Hello, Can you please provide the summary of your project and also the accuracy results? Thank you.

erko commented 6 years ago

Hi @pGit1, Added poster files describing about my solution (were hanged at NIPS 2017 workshop). Very soon will be posted NIPS competitions track book here on github. There my and other solutions were described more detailed.

Particularly about my solution, in two words: very simple defense tactic using median filtering fed to only adversarially trained models. In my submitted solution I've used ensemble of adversarially pretrained models listed in this repo: https://github.com/tensorflow/models/. So, doesn't even need additional expensive trainings if you already adversarially trained models. Final results: 4th place on leaderboard.

This solution are applicable to big images like in ImageNet and not general. So, problem is far from solved.

pGit1 commented 6 years ago

THANK YOU!

pGit1 commented 6 years ago

@erko

Could you provide a link to this:

Very soon will be posted NIPS competitions track book here on github

Thank you! Great job by the way

erko commented 6 years ago

@pGit1 It up to NIPS organizers, they asked to not share it before their release.. So, I'm waiting them too.