thanh-isu / unrestricted-adversarial-examples

Contest Proposal and infrastructure for the Unrestricted Adversarial Examples Challenge

Unrestricted Adversarial Examples Challenge

In the Unrestricted Adversarial Examples Challenge, attackers submit arbitrary adversarial inputs, and defenders are expected to assign low confidence to difficult inputs while retaining high confidence and accuracy on a clean, unambiguous test set. You can learn more about the motivation and structure of the contest in our recent paper

This repository contains code for the warm-up to the challenge, as well as the public proposal for the contest. We are currently accepting defenses for the warm-up.

Leaderboard for the warm-up to the contest

We include three attacks in the warm-up to the contest:

• 1000 Linfinity-ball adversarial examples generated by SPSA
• 1000 spatial adversarial examples (via grid search)
• 100 L2-ball adversarial examples generated by the Boundary attack

The top few distinct models for each dataset are shown below. You can see all submissions in the full scoreboard.

Two-Class MNIST dataset

Defense	Submitted by	Clean data	Spatial grid attack	SPSA attack	Boundary attack	Submission Date
MadryPGD LeNet Baseline	Google Brain	100.0%	0%	19.6%	0%	Sept 14th, 2018
Undefended LeNet Baseline	Google Brain	100.0%	0%	0%	0%	Sept 14th, 2018

All percentages above correspond to the model's accuracy at 80% coverage.

Bird or Bicycle dataset

Defense	Submitted by	Clean data	Common corruptions	Spatial grid attack	SPSA attack	Boundary attack	Submission Date
Keras ResNet (trained on ImageNet)	Google Brain	100.0%	99.2%	92.2%	1.6%	4.0%	Sept 29th, 2018
Pytorch ResNet (trained on bird-or-bicycle extras)	Google Brain	98.8%	74.6%	49.5%	2.5%	8.0%	Oct 1st, 2018

All percentages above correspond to the model's accuracy at 80% coverage.

Submitting a defense for the warm-up

The warm-up before the contest is currently underway and is accepting submissions. If you have additional questions, feel free to submit a new GitHub issue with the "question" tag and we will respond shortly.

The contest

The contest phase will begin after the warm-up attacks have been conclusively solved. We have published the contest proposal and are soliciting feedback from the community.

Paper

You can learn more about the motivation and structure of the contest in our recent paper:

Unrestricted Adversarial Examples
Tom B. Brown, Nicholas Carlini, Chiyuan Zhang, Catherine Olsson, Paul Christiano and Ian Goodfellow
Arxiv preprint

@article{unrestricted_advex_2018,
  title = {Unrestricted Adversarial Examples},
  author = {{Brown}, T.~B. and {Carlini}, N. and {Zhang}, C. and {Olsson}, C. and 
	  {Christiano}, P. and {Goodfellow}, I.},
  journal={arXiv preprint arXiv:1809.08352},
  year={2018}
}