ModelGuard: Information-Theoretic Defense Against Model Extraction Attacks


Minxue Tang and Anna Dai, Duke University; Louis DiValentin, Aolin Ding, and Amin Hass, Accenture; Neil Zhenqiang Gong, Yiran Chen, and Hai "Helen" Li, Duke University


Malicious utilization of a query interface can compromise the confidentiality of ML-as-a-Service (MLaaS) systems via model extraction attacks. Previous studies have proposed to perturb the predictions of the MLaaS system as a defense against model extraction attacks. However, existing prediction perturbation methods suffer from a poor privacy-utility balance and cannot effectively defend against the latest adaptive model extraction attacks. In this paper, we propose a novel prediction perturbation defense named ModelGuard, which aims at defending against adaptive model extraction attacks while maintaining a high utility of the protected system. We develop a general optimization problem that considers different kinds of model extraction attacks, and ModelGuard provides an information-theoretic defense to efficiently solve the optimization problem and achieve resistance against adaptive attacks. Experiments show that ModelGuard attains significantly better defensive performance against adaptive attacks with less loss of utility compared to previous defenses.

Open Access Media

USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.