Key Attack Strategies Against Black-Box DNNs

Yassine Hmamouche (1, 2) , Yehya Nasser (1, 3) , Amer Baghdadi (1, 3) , Marc-Oliver Pahl (1, 4)

1 IMT Atlantique - MEE - Département Mathematical and Electrical Engineering
2 Lab-STICC_COSYDE - Equipe Communication System Design
3 Lab-STICC_2AI - Equipe Algorithm Architecture Interactions
4 OCIF - Objets communicants pour l'Internet du futur

In this paper, we examined to what extent and under what settings the confidentiality and integrity of black-box DNNs—which are the most challenging setup of DNNs—can be threatened. In this way, we proposed a comprehensive taxonomy of the key strategies developed in the literature to attack black-box DNNs.We believe that a coherent classification incorporating all key aspects is needed to organise the body of knowledge on research and methodologies for understanding and securing black-box DNNs.

Keywords: Deep neural network; side-channel attack

Domains: Engineering Sciences [physics]

Yassine Hmamouche, Yehya Nasser, Amer Baghdadi, Marc-Oliver Pahl. Key Attack Strategies Against Black-Box DNNs. GDR-SOC2, Jun 2022, Strasbourg, France. , GDR SOC2. ⟨hal-03690454⟩