Exploring Image Reconstruction Attack in Deep Learning Computation Offloading

Abstract

Deep learning (DL) computation offloading is commonly adopted to enable the use of computation-intensive DL techniques on resource-constrained devices. However, sending private user data to an external server raises a serious privacy concern. In this paper, we introduce a privacy-invading input reconstruction method which utilizes intermediate data of the DL computation pipeline. In doing so, we first define a Peak Signal-to-Noise Ratio (PSNR)-based metric for assessing input reconstruction quality. Then, we simulate a privacy attack on diverse DL models to find out the relationship between DL model structures and performance of privacy attacks. Finally, we provide several insights on DL model structure design to prevent reconstruction-based privacy attacks: using skip-connection, making model deeper, including various DL operations such as inception module.