Flexible cross-modal steganography via implicit representations

Abstract

We present INRSteg, an innovative lossless steganography framework based on a novel data form, Implicit Neural Representations (INR), that is modal-agnostic. Our framework effectively hides multiple data without altering the original INR ensuring high-quality stego data. The neural representations of secret data are first concatenated to have independent paths that do not overlap. Then weight freezing techniques are applied to the diagonal blocks of the concatenated network's weight matrices to preserve the weights of secret data while the additional free weights in the off-diagonal blocks of weight matrices are fitted to the cover data. Our framework can perform unexplored cross-modal steganography for various modalities including image, audio, video, and 3D shapes, and it achieves state-of-the-art performance compared to previous intra-modal steganographic methods.