Transformer Driven Self-Attention Framework for High Fidelity Color Image Steganography

Abstract

A significant issue in digital steganography is concealing color images within natural photographs, when the objective is to convey a large quantity of hidden information without sacrificing visual quality.  Certain conventional techniques that rely on the spatial or frequency domains typically leave behind noticeable artifacts and are readily interfered with by filtering, noise, or compression.  Some transformer-based models have recently demonstrated exceptional ability to capture intricate linkages and long-range interdependence across data channels.  Building on this, we have presented a transformer-powered framework for color image concealment that learns the best embedding patterns across RGB channels by using self-attention. In comparison to both conventional techniques and previous deep learning methods, our results demonstrate that this methodology provides considerably greater imperceptibility and recovery, indicating transformers as a promising avenue for enhanced image steganography.