Yucheng Shi and Siyu Wang won the fourth place in both Untargeted Attack Track and Targeted Attack Track of NIPS 2018 Adversarial Vision Challenge. In this competition, more than 300 models and attacks have been submitted by more than 400 teams from all over the world. Our team, which is named ‘Degurechaff’, is the only team in China that has entered Top-5 in both attack tracks.

NIPS 2018 Adversarial Vision Challenge is designed to facilitate measurable progress towards robust machine vision models and more generally applicable adversarial attacks. There are three tracks included: Robust Model Track, Untargeted Attacks Track and Targeted Attacks Track. Different from previous competitions, like NIPS 2017 or CAAD, NIPS 2018 competition uses median l2 distance to measure the quality of adversarial examples generated. In addition, all submissions are continuously pitted against each other on a fixed set of samples to encourage a co-evolution of robust models and better adversarial attacks, which guarantees the fairness of evaluation.

In this competition, we propose a novel iterative black-box attack that can effectively improve the transferability and diversity of adversarial examples. Moreover, we for the first time design an optimization method that can squeeze out redundant adversarial noises. Paper and technical details about new attack methods will be released soon.

Leaderboard of final results for Untargeted Attack and Targeted Attack track:

https://www.crowdai.org/challenges/nips-2018-adversarial-vision-challenge-untargeted-attack-track/leaderboards https://www.crowdai.org/challenges/nips-2018-adversarial-vision-challenge-targeted-attack-track/leaderboards