Can Go AIs be adversarially robust? (arxiv.org)

Recent research has revealed significant vulnerabilities in superhuman Go AIs, which can be defeated through simple adversarial strategies, particularly cyclic attacks. The study investigated the effectiveness of various countermeasures, including adversarial training on curated positions, iterated adversarial training, and adjustments to the AI’s network architecture. While some defenses provided limited protection against known attacks, they ultimately fell short against newly trained adversarial strategies. This highlights a critical challenge in AI robustness: even advanced systems struggle to withstand evolving threats in an inherently adversarial environment. The findings are significant for the AI/ML community, emphasizing that enhancing adversarial robustness is a complex task that necessitates more efficient defense generalization and greater training diversity. As researchers forge ahead with AI developments, the study underlines the necessity for continual adaptation to new forms of attacks. These insights not only raise caution about the readiness of current AI systems but also signal the importance of ongoing exploration in adversarial training methodologies to develop more resilient AI frameworks. For those interested, interactive attack demonstrations and the associated codebase are available for further examination.