Title

Towards Stealing Deep Neural Networks on Mobile Devices

Document Type

Book Chapter

Publication Date

11-4-2021

Department

Department of Computer Science

Abstract

Recently, deep neural networks (DNN) are increasingly deployed on mobile computing devices. Compared to the traditional cloud-based DNN services, the on-device DNN provides immediate responses without relying on network availability or bandwidth and can boost security and privacy by preventing users’ data from transferring over the untrusted communication channels or cloud servers. However, deploying DNN models on the mobile devices introduces new attack vectors on the models. Previous studies have shown that the DNN models are prone to model stealing attacks in the cloud setting, by which the attackers can steal the DNN models accurately. In this work, for the first time, we study the model stealing attacks on the deep neural networks running in the mobile devices, by interacting with mobile applications. Our experimental results on various datasets confirm the feasibility of stealing DNN models in mobile devices with high accuracy and small overhead.

Publisher's Statement

© ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2021. Publisher’s version of record: https://doi.org/10.1007/978-3-030-90022-9_27

Publication Title

Security and Privacy in Communication Networks

Share

COinS