2026-05-19T10:55:07Z https://ipsj.ixsq.nii.ac.jp/oai

oai:ipsj.ixsq.nii.ac.jp:00232730 2025-01-19T10:20:24Z 1164:4619:11539:11552

閾値操作による公平な分類アルゴリズムの公平性に対する毒攻撃 Poisoning Attack on Fairness of Fair Classification Algorithm through Threshold Control. ダイ, ショウテン 秋本, 洋平 佐久間, 淳 福地, 一斗 Shengtian, Dai Youhei, Akimoto Jun, Sakuma Kazuto, Fukuchi The ethical issues of artificial intelligence have become more severe as machine learning is widely used in several fields. Recent developments in machine learning enable machine learning algorithms to mitigate the ethical issue of fairness, a problem of discriminating output against sensitive attributes, including gender and race. However, a poisoning attack, originally designed to harm the accuracy of models, can also introduce unfair bias into models. Our research investigated the attack of worsening fairness of the fair learning models, finding how the fair learning models behave under the fairness attack. Specifically, we construct an attack strategy, TAF, targeting the fairness of the fair learning model by controlling the thresholds involved in the model and elucidate its behavior. The experimental results demonstrate that TAF does more harm to the fairness of the fair learning model than the attack methods proposed in existing studies. The ethical issues of artificial intelligence have become more severe as machine learning is widely used in several fields. Recent developments in machine learning enable machine learning algorithms to mitigate the ethical issue of fairness, a problem of discriminating output against sensitive attributes, including gender and race. However, a poisoning attack, originally designed to harm the accuracy of models, can also introduce unfair bias into models. Our research investigated the attack of worsening fairness of the fair learning models, finding how the fair learning models behave under the fairness attack. Specifically, we construct an attack strategy, TAF, targeting the fairness of the fair learning model by controlling the thresholds involved in the model and elucidate its behavior. The experimental results demonstrate that TAF does more harm to the fairness of the fair learning model than the attack methods proposed in existing studies. 情報処理学会 2024-02-25 eng technical report https://ipsj.ixsq.nii.ac.jp/records/232730 2188-8701 AA11131797 研究報告コンピュータビジョンとイメージメディア（CVIM） 2024-CVIM-237 39 1 8 https://ipsj.ixsq.nii.ac.jp/record/232730/files/IPSJ-CVIM24237039.pdf application/pdf 1.1 MB