Can Go AIs be adversarially robust?
The paper investigates the adversarial robustness of superhuman Go AIs against simple adversarial strategies. Despite implementing various countermeasures, the study finds that none of the defenses are fully effective against newly trained adversaries. The results emphasize the challenges in developing robust AI systems, even in favorable settings like Go.
- ▪Superhuman Go AIs can be defeated by simple adversarial strategies, particularly cyclic attacks.
- ▪The study tests three defenses, including adversarial training and changes to network architecture.
- ▪None of the defenses withstand attacks from freshly trained adversaries, highlighting gaps in efficient generalization and training diversity.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2406.12843 (cs) [Submitted on 18 Jun 2024 (v1), last revised 14 Jan 2025 (this version, v3)] Title:Can Go AIs be adversarially robust? Authors:Tom Tseng, Euan McLean, Kellin Pelrine, Tony T. Wang, Adam Gleave View a PDF of the paper titled Can Go AIs be adversarially robust?, by Tom Tseng and 4 other authors View PDF Abstract:Prior work found that superhuman Go AIs can be defeated by simple adversarial strategies, especially "cyclic" attacks. In this paper, we study whether adding natural countermeasures can achieve robustness in Go, a favorable domain for robustness since it benefits from incredible average-case capability and a narrow, innately adversarial setting.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
