{"metadata":{"_oai":{"id":"oai:ipsj.ixsq.nii.ac.jp:00214434","sets":["6164:6165:6462:10749"]},"path":["10749"],"owner":"44499","recid":"214434","title":["差分プライベートなベイジアンニューラルネットワークのプライバシーリスク"],"pubdate":{"attribute_name":"公開日","attribute_value":"2021-10-19"},"_buckets":{"deposit":"06f5217f-1e3a-4dc8-91cc-dd3ce806372a"},"_deposit":{"id":"214434","pid":{"type":"depid","value":"214434","revision_id":0},"owners":[44499],"status":"published","created_by":44499},"item_title":"差分プライベートなベイジアンニューラルネットワークのプライバシーリスク","author_link":["550549","550551","550552","550545","550548","550547","550550","550546"],"item_titles":{"attribute_name":"タイトル","attribute_value_mlt":[{"subitem_title":"差分プライベートなベイジアンニューラルネットワークのプライバシーリスク"},{"subitem_title":"Privacy Risk of Differentially Private Bayesian Neural Network","subitem_title_language":"en"}]},"item_keyword":{"attribute_name":"キーワード","attribute_value_mlt":[{"subitem_subject":"差分プライバシー，ベイジアンニューラルネットワーク，深層学習","subitem_subject_scheme":"Other"}]},"item_type_id":"18","publish_date":"2021-10-19","item_language":{"attribute_name":"言語","attribute_value_mlt":[{"subitem_language":"jpn"}]},"item_18_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":"NTT社会情報研究所"},{"subitem_text_value":"NTT社会情報研究所"},{"subitem_text_value":"NTT社会情報研究所"}]},"item_18_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":"NTT Social Informatics Laboratories","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"}]},"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/214434/files/IPSJCSS2021034.pdf","label":"IPSJCSS2021034.pdf"},"date":[{"dateType":"Available","dateValue":"2023-10-19"}],"format":"application/pdf","billing":["billing_file"],"filename":"IPSJCSS2021034.pdf","filesize":[{"value":"1.1 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":"30"},{"tax":["include_tax"],"price":"0","billingrole":"46"},{"tax":["include_tax"],"price":"0","billingrole":"44"}],"accessrole":"open_date","version_id":"843200d1-a5b5-48fa-8dbd-2018e4684313","displaytype":"detail","licensetype":"license_note","license_note":"Copyright (c) 2021 by the Information Processing Society of Japan"}]},"item_18_creator_5":{"attribute_name":"著者名","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"芝原, 俊樹"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"三浦, 尭之"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"紀伊, 真昇"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"市川, 敦謙"}],"nameIdentifiers":[{}]}]},"item_18_creator_6":{"attribute_name":"著者名(英)","attribute_type":"creator","attribute_value_mlt":[{"creatorNames":[{"creatorName":"Toshiki, Shibahara","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Takayuki, Miura","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Masanobu, Kii","creatorNameLang":"en"}],"nameIdentifiers":[{}]},{"creatorNames":[{"creatorName":"Atsunori, Ichikawa","creatorNameLang":"en"}],"nameIdentifiers":[{}]}]},"item_resource_type":{"attribute_name":"資源タイプ","attribute_value_mlt":[{"resourceuri":"http://purl.org/coar/resource_type/c_5794","resourcetype":"conference paper"}]},"item_18_description_7":{"attribute_name":"論文抄録","attribute_value_mlt":[{"subitem_description":"ベイジアンニューラルネットワーク（NN）は，予測値の不確実性も出力できるため，deep neural network (DNN) の誤識別が深刻な影響を及ぼすタスクで注目されている．特に，Monte Calro (MC) dropout を用いると，NN の構造を変えずに予測時の使い方を変えるだけで，決定的な NN をベイジアン NN として使うことができる．このとき，ベイジアン NN では，より多くの情報を出力するため，教師データの情報が漏洩するリスクも高まる可能性があるが，詳細な調査は行われていない．そこで，本稿では，DNN を決定的な NN として使用した場合と， MC dropout を適用してベイジアン NN として使用した場合で，プライバシーリスクがどの程度異なるかを評価する．プライバシーリスクの評価は，ベイジアン NN の利用シーンを想定して定義した攻撃モデルに従い，ベイジアン NN の出力に基づいてリスクを評価する提案手法を用いて行った．CIFAR-10 と CNN を使用した実験で，ベイジアン NN は決定的な NN よりプライバシーリスクが高いこと，ベイジアン NN の出力を量子化することで有用性を維持しつつプライバシーリスクを低減できる可能性が高いことを示した．","subitem_description_type":"Other"}]},"item_18_description_8":{"attribute_name":"論文抄録(英)","attribute_value_mlt":[{"subitem_description":"Bayesian neural networks (NNs) that output uncertainties of predictions have attracted attention in a critical situation. Especially, Monte Calro (MC) dropout enables us to use deterministic NNs as Bayesian NNs without changing NN architectures. Since Bayesian NNs output more information than deterministic NNs, Bayesian NNs potentially have higher privacy risks. However, whether applying MC dropout increases privacy risks of deep NNs (DNNs) or not has not been studied. Therefore, we compare privacy risks of deterministic NNs to those of Bayesian NNs. For the comparison, we define an attack model against Bayesian NNs and design a method for evaluating privacy risks of Bayesian NNs. Our experiments using CIFAR-10 and CNN show that Bayesian NNs have higher privacy risks than deterministic NNs and that privacy risks can be decreased without degrading the utility by quantizing outputs of Bayesian NNs.","subitem_description_type":"Other"}]},"item_18_biblio_info_10":{"attribute_name":"書誌情報","attribute_value_mlt":[{"bibliographicPageEnd":"252","bibliographic_titles":[{"bibliographic_title":"コンピュータセキュリティシンポジウム2021論文集"}],"bibliographicPageStart":"245","bibliographicIssueDates":{"bibliographicIssueDate":"2021-10-19","bibliographicIssueDateType":"Issued"}}]},"relation_version_is_last":true,"weko_creator_id":"44499"},"id":214434,"updated":"2025-01-19T16:38:03.152806+00:00","links":{},"created":"2025-01-19T01:15:15.134932+00:00"}