A statistical threshold for adversarial classification in laplace mechanisms

Unsal, Ayse; Önen, Melek
ITW 2021, IEEE Information Theory Workshop, 17-21 October 2021, Kanazawa, Japan (Virtual Event)

This paper studies the statistical characterization of detecting an adversary who wants to harm some computation such as machine learning models or aggregation by altering the output of a differentially private mechanism in addition to discovering some information about the underlying dataset. An adversary who is able to modify the published information from a differentially private mechanism aims to maximize the possible damage to the system while remaining undetected. We present a trade-off between the privacy parameter of the system, the sensitivity and the attacker’s advantage (the bias) through determining the threshold for the best critical region of the hypothesis testing problem for deciding whether or not the adversary’s attack is detected. Such tradeoffs are provided for Laplace mechanisms using one-sided and two-sided hypothesis tests. Corresponding error probabilities are analytically derived and ROC curves are presented for various levels of the sensitivity, the absolute mean of the attack and the privacy parameter. Subsequently, we provide an interval for the bias induced by the adversary so that the defender detects the attack. Finally, we adapt the Kullback-Leibler differential privacy to adversarial classification. A full version of this paper is accessible at: https://arxiv.org/abs/2105.05610.

Sécurité numérique
Eurecom Ref:
© 2021 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

PERMALINK : https://www.eurecom.fr/publication/6488