{"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00225724","sets":["1164:4088:11174:11249"]},"path":["11249"],"owner":"44499","recid":"225724","title":["検証可能な機械学習モデルの利用に向けた一検討：画像加工への頑健性に基づくモデルの同一性検証手法の提案"],"pubdate":{"attribute_name":"公開日","attribute_value":"2023-05-04"},"_buckets":{"deposit":"30114d7f-d35e-4b8f-bd36-5cdd74192ec0"},"_deposit":{"id":"225724","pid":{"type":"depid","value":"225724","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"検証可能な機械学習モデルの利用に向けた一検討：画像加工への頑健性に基づくモデルの同一性検証手法の提案","author_link":["597889","597891","597890","597893","597888","597892","597894"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"検証可能な機械学習モデルの利用に向けた一検討：画像加工への頑健性に基づくモデルの同一性検証手法の提案"},{"subitem_title":"Consideration toward the Use of Verifiable Machine Learning Models: Proposal of the Method to Verify Model's Identity Based on Robustness to Image Processing","subitem_title_language":"en"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"CSEC, セキュリティ","subitem_subject_scheme":"Other"}]},"item_type_id":"4","publish_date":"2023-05-04","item_4_text_3":{"attribute_name":"著者所属","attribute_value_mlt":[{"subitem_text_value":"NTT社会情報研究所"},{"subitem_text_value":"静岡大学"},{"subitem_text_value":"NTT社会情報研究所"},{"subitem_text_value":"NTT社会情報研究所"},{"subitem_text_value":"静岡大学"},{"subitem_text_value":"静岡大学"},{"subitem_text_value":"静岡大学"}]},"item_4_text_4":{"attribute_name":"著者所属(英)","attribute_value_mlt":[{"subitem_text_value":"NTT Social Informatics Laboratories","subitem_text_language":"en"},{"subitem_text_value":"Shizuoka University","subitem_text_language":"en"},{"subitem_text_value":"NTT Social Informatics Laboratories","subitem_text_language":"en"},{"subitem_text_value":"NTT Social Informatics Laboratories","subitem_text_language":"en"},{"subitem_text_value":"Shizuoka University","subitem_text_language":"en"},{"subitem_text_value":"Shizuoka University","subitem_text_language":"en"},{"subitem_text_value":"Shizuoka University","subitem_text_language":"en"}]},"item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_publisher":{"attribute_name":"出版者","attribute_value_mlt":[{"subitem_publisher":"情報処理学会","subitem_publisher_language":"ja"}]},"publish_status":"0","weko_shared_id":-1,"item_file_price":{"attribute_name":"Billing file","attribute_type":"file","attribute_value_mlt":[{"url":{"url":"https://ipsj.ixsq.nii.ac.jp/record/225724/files/IPSJ-IOT23061016.pdf","label":"IPSJ-IOT23061016.pdf"},"date":[{"dateType":"Available","dateValue":"2025-05-04"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJ-IOT23061016.pdf","filesize":[{"value":"1.0 MB"}],"mimetype":"application/pdf","priceinfo":[{"tax":["include_tax"],"price":"660","billingrole":"5"},{"tax":["include_tax"],"price":"330","billingrole":"6"},{"tax":["include_tax"],"price":"0","billingrole":"43"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"040109ca-0b09-47d6-8f6e-faec3c47d9d9","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2023 by the Information Processing Society of Japan"}]},"item_4_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"桐淵, 直人"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"佐藤, 佑哉"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"鈴木, 亮平"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"芦澤, 奈実"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"大木, 哲史"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"峰野, 博史"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"西垣, 正勝"}],"nameIdentifiers":[{}]}]},"item_4_source_id_9":{"attribute_name":"書誌レコードID","attribute_value_mlt":[{"subitem_source_identifier":"AA12326962","subitem_source_identifier_type":"NCID"}]},"item_4_textarea_12":{"attribute_name":"Notice","attribute_value_mlt":[{"subitem_textarea_value":"SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc."}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_18gh","resourcetype":"technical report"}]},"item_4_source_id_11":{"attribute_name":"ISSN","attribute_value_mlt":[{"subitem_source_identifier":"2188-8787","subitem_source_identifier_type":"ISSN"}]},"item_4_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"他者が提供する機械学習モデル（以降，AI と呼ぶ）を利用する際に，安心した AI 利用の促進に向けて，利用者が期待する AI かを検証可能にする検討を行う．その一環として，事前に取得した期待する AI の情報と利用する AI との同一性を検証する手法を提案する．提案手法は，画像を扱う AI を対象に，加工した画像に対する推論の頑健性をフィンガープリントとすることで，AI 内部のパラメータを直接参照せずに AI の特性を調べ，複数枚の画像に対する特性を集めて統計的に処理することで AI の同一性を検証する．利用者が検証したい AI に期待する事柄は状況に応じて様々であり，そのすべてに対応するフレームワークを構築することが本研究の目的である．その内，本稿では，利用者の大きな関心事として AI がどのように作成されたかという観点で，AI を作成する際の重要な要素である学習に用いたデータセットに着目する．実験を通じて，重複のない異なるデータセットで学習した AI を提案手法により識別できることを示す．また，学習における初期のランダムなパラメータなどの差異によって AI 自体が完全に同一でなくても，同じデータセットで学習すれば同一の AI だと判定できることも確認する．提案手法の応用例として，AI の作成者が主張するデータセットで学習したか否かを利用者が検証することで，作成者による虚偽の主張の発見や，特性に応じた適切な AI の利用判断に役立つことが期待される．","subitem_description_type":"Other"}]},"item_4_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"When using machine learning models (hereafter referred to as AI) provided by others, this study aims to make it possible to verify whether users expect AI in order to promote the use of AI with peace of mind. As part of this, we propose a method verifying the identity of the expected AI information obtained beforehand with the AI to be used. The proposed method examines the characteristics of AI by making the robustness of inference to processed images as fingerprint for AI handling images without directly referring to the internal parameters of AI. Furthermore, it verifies the identity of AI by collecting and statistically processing the characteristics against multiple images. What users want to verify about AI varies depending on the situation. This study aims to build a framework for all of them. Among them, this paper focuses on the dataset used for learning, which is an essential factor in creating AI, from the perspective of how AI was created as a primary concern of users. Through experiments, we show that the proposed method can identify learned AIs in different datasets without duplication, even if the AIs' architecture is the same. We also confirm that even if the AI is not entirely the same due to differences such as initial random parameters in learning, it can be determined to be the same by learning on the same dataset. As an application of the proposed method, verifying whether the user has learned from the dataset claimed by the AI's creator will help detect false claims by the creator and for judging the appropriate use of AI according to the characteristics.","subitem_description_type":"Other"}]},"item_4_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"7","bibliographic_titles":[{"bibliographic_title":"研究報告インターネットと運用技術（IOT）"}],"bibliographicPageStart":"1","bibliographicIssueDates":{"bibliographicIssueDate":"2023-05-04","bibliographicIssueDateType":"Issued"},"bibliographicIssueNumber":"16","bibliographicVolumeNumber":"2023-IOT-61"}]},"relation_version_is_last":true,"weko_creator_id":"44499"},"updated":"2025-01-19T12:42:07.935539+00:00","created":"2025-01-19T01:25:13.678630+00:00","links":{},"id":225724}