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.
The latest submission by Chongli Qin et al has claimed to solve the warm-up to the challenge. We are verifying the submission with our advisory board, and preparing to launch the full-fledged version of 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.
Defense | Submitted by | Clean data | Spatial grid attack | SPSA attack | Boundary attack | Submission Date | Open Source |
---|---|---|---|---|---|---|---|
MadryPGD LeNet Baseline | Google Brain | 100.0% | 0% | 19.6% | 0% | Sept 14th, 2018 | Yes |
Undefended LeNet Baseline | Google Brain | 100.0% | 0% | 0% | 0% | Sept 14th, 2018 | Yes |
All percentages above correspond to the model's accuracy at 80% coverage.
Defense | Submitted by | Clean data | Common corruptions | Spatial grid attack | SPSA attack | Boundary attack | Submission Date | Open Source |
---|---|---|---|---|---|---|---|---|
LLR_ADV_TRAIN | Chongli Qin & Jonathan Uesato | 100.0% | 100.0% | 100.0% | 100.0% | 100.0% | Dec 14th, 2019 | Yes |
TRADESv2 | Hongyang Zhang (CMU) & Xin Li (Lehigh Univ.) | 100.0% | 100.0% | 99.5% | 100.0% | 95.0% | Jan 17th, 2019 | No |
Keras ResNet (trained on ImageNet) |
Google Brain | 100.0% | 99.2% | 92.2% | 1.6% | 4.0% | Sept 29th, 2018 | Yes |
Pytorch ResNet (trained on bird-or-bicycle extras) |
Google Brain | 98.8% | 74.6% | 49.5% | 2.5% | 8.0% | Oct 1st, 2018 | Yes |
All percentages above correspond to the model's accuracy at 80% coverage.
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 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.
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}
}