ウダ　リュウヤ   宇田　隆哉    所属   コンピュータサイエンス学部 コンピュータサイエンス学科    職種   准教授
言語種別	英語
発行・発表の年月	2024/02
形態種別	国際会議論文
査読	査読あり
標題	Invisible Cloak to AI Recognition from All Horizontal Directions by Adversarial Patch
執筆形態	共著
掲載誌名	2024 18th International Conference on Ubiquitous Information Management and Communication (IMCOM)
掲載区分	国外
出版社・発行元	IEEE
総ページ数	7
担当区分	最終著者
著者・共著者	Takumi Imaeda and Ryuya Uda
概要	Humans have a right not to be recognized by AI without permission while surveillance cameras and drones are coming to recognize humans automatically. The adversarial patch is one of the technologies that prevents objects from being recognized by AI. The patch was applied to a board to work in the real world. Moreover, it was improved to be a T-shirt with wrinkle resistance. Furthermore, it was also improved to have shooting angle resistance. However, all patches in the existing research do not have enough shooting resistance from all angles, especially from the side and from the back. Therefore, we proposed a cloak that had shooting resistance from all angles, but a pattern in a patch was sometimes misrecognized as a human. Hence, in this paper, we improved our cloak by changing the initial image to create adversarial patches. We also evaluated the recognition by AI. As a result, every pattern in a patch was no longer misrecognized as a human. On the contrary, the human detection rate increased to about 26%, while in the previous cloak it was 1.2%. We considered the reason for the increase to be the color of the initial image, and found a better color than the original image. We believe that a cyan initial image will make our cloak completely invisible to the AI recognition from all horizontal directions.
researchmap用URL	https://doi.org/10.1109/IMCOM60618.2024.10418360