Counterfactual Image Generation using GAN for Fairness

• Item type: SIG Technical Reports(1)
• 公開日: 2021-02-25
• タイトル: Counterfactual Image Generation using GAN for Fairness
• タイトル / 言語: en
• タイトル: Counterfactual Image Generation using GAN for Fairness
• 言語: eng
• キーワード / 主題Scheme: Other
• キーワード / 主題: セッション1-1
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_18gh
• 資源タイプ: technical report
• 著者所属: Presently with Computational Science, System Informatics, Kobe University Graduate School
• 著者所属: Presently with Graduate School of Engineering Science, Osaka University
• 著者所属: Presently with Faculty of Business Administration, Osaka Gakuin University
• 著者所属(英): en
• 著者所属(英): Presently with Computational Science, System Informatics, Kobe University Graduate School
• 著者所属(英): en
• 著者所属(英): Presently with Graduate School of Engineering Science, Osaka University
• 著者所属(英): en
• 著者所属(英): Presently with Faculty of Business Administration, Osaka Gakuin University
• 著者名: Koki, Wataoka; Takashi, Matsubara; Kuniaki, Uehara
• 著者名(英): Koki, Wataoka; Takashi, Matsubara; Kuniaki, Uehara
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: Computer vision systems have made significant improvements and been used in a variety of situations. For a practical use, we need to prevent the systems from making unfair decisions for certain individuals. In this sense, the systems have to eliminate the difference between decision makings on the real world and the counterfactual world where users would have different sensitive attributes (e.g., gender and race). In this study, we propose a framework for counterfactual image generation named Causality with Unobserved Variables using Generative Adversarial Networks (CUV-GAN). CUV-GAN can generate counterfactual images as the results of the intervention in the images' attributes and improve the fairness of an image classifier by being trained with generated images as data augmentation.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: Computer vision systems have made significant improvements and been used in a variety of situations. For a practical use, we need to prevent the systems from making unfair decisions for certain individuals. In this sense, the systems have to eliminate the difference between decision makings on the real world and the counterfactual world where users would have different sensitive attributes (e.g., gender and race). In this study, we propose a framework for counterfactual image generation named Causality with Unobserved Variables using Generative Adversarial Networks (CUV-GAN). CUV-GAN can generate counterfactual images as the results of the intervention in the images' attributes and improve the fairness of an image classifier by being trained with generated images as data augmentation.
• 書誌レコードID / 収録物識別子タイプ: NCID
• 書誌レコードID / 収録物識別子: AA11131797
• 書誌情報: 研究報告コンピュータビジョンとイメージメディア（CVIM）; 巻 2021-CVIM-225, 号 4, p. 1-7, 発行日 2021-02-25
• ISSN / 収録物識別子タイプ: ISSN
• ISSN / 収録物識別子: 2188-8701
• Notice: SIG Technical Reports are nonrefereed and hence may later appear in any journals, conferences, symposia, etc.
• 出版者 / 言語: ja
• 出版者: 情報処理学会