[4N2-OS-26a] OS-26 (1)

[4N2-OS-26a-01] Vulnerability to Adversarial Attacks caused by Fairness

〇Koki Wataoka¹, Takashi Matsubara¹, Kuniaki Uehara¹ (1. Kobe University)

Keywords:Fairness, Adversarial Attacks, Machine Learning, Classifier

It is essential to ensure that classifiers are fair to gender and age. Hence, many methods have been proposed to make classifiers fair. However, there is no discussion about the security of fair classifiers. In order to secure machine learning, we must consider adversarial attacks that degrade the accuracy of classifiers. In this paper, we demonstrate that fair classifiers are vulnerable to adversarial attacks. Our experiments have shown that fair classifiers are less robust against adversarial attacks than usual classifiers, and hence show worse classification accuracy and fairness performances.