Hiding Data in Images Using Spectral Filtering and Deep Neural Networks
Hadar Shalev, Pe’erly Setter, Ruth Kimchi, Hagit Zabrodsky Hel-Or · 2020
Privacy preserving and data security is one of the major concerns of our world. In this study we deal with one aspect, namely shoulder surfing, where sensitive information may be accessed by looking at the display or keyboard of the user. We propose a method of hiding information in color images so that it can not be perceived by the naked eye and requires a spectral filter to be seen. We term such images Spectral Hiding Images. In this work we developed a system which can automatically generate Spectral Hiding Images. We focus on a basic class of images containing a single numeric digit. We train three deep networks to determine the salient digit in an image, determine the hidden and masked digits in a spectral hiding image and to generate diverse spectral hiding digit images. Mass producing such Spectral Hiding Images using our system will allow for screen content hiding and password hiding. Additionally we show that several Gestalt principles of human perception, are expressed in the trained networks' behavior.