Explainable AI based Adversarial Examples and its Evaluation
Ryo Kumagai, Shu Takemoto, Yusuke Nozaki, 雅弥 吉川 · 2023
With the advancement of AI technology, AI can be applied to various fields. Currently, AI is used for vital decision such as medical diagnosis. Therefore, the accountability for the decision by AI has become more important. The explainable AI (XAI) is a key technology to fulfill the accountability. While AI is spreading rapidly, the vulnerability of AI system is pointed out. The adversarial examples (AE), which causes wrong decisions by AI, is one of the terrible attacks for AI. For using AI safely, AE has to be investigated thoroughly. In order to clarify mechanism and affect of AE, this paper proposes the generating method for AE which is based on XAI. Experiments prove the proposed method is superior to previous AEs.