StegaNeRV: Video Steganography Using Implicit Neural Representation

Abstract

Numerous studies have recently advanced the state-of-the art for representing videos through an implicit neural network (INR). As these models become increasingly ubiquitous, there is a growing demand for concealing data within INR reconstructed videos such as for storing content metadata and sensitive licensing information. In this paper, we explore a new space in video steganography, hiding a distinct image within each RGB frame output by an INR. We propose a joint training strategy of a U-Net based steganographic decoder with an INR model for video. Experimental results show that hidden images can be embedded and subsequently reconstructed with high fidelity while preserving the quality of the cover frames. Furthermore we demonstrate that by introducing an attention module which emphasizes hiding within the edges and rich texture patches in the cover frame, secret images can be reconstructed with superior quality and can also be concealed at greater resolutions.