A cross-domain embedding cost learning joint FFT for security steganography

Abstract

Recent advancements in image steganography demonstrate that reasonable probability maps generated by minimum embedding cost learning through adversarial training can effectively improve the security performance of steganography. Existing embedding cost learning based steganography methods primarily rely on the generator to extract structural features in the image spatial domain, neglecting high frequency information in the frequency domain, which restricts the performance of the model. To address this gap, we propose a minimum embedding cost learning network based on a cross-domain feature fusion, not only extracting the spatial domain information, but also identifying the features in frequency information, aiming to generate effective probability maps for steganography. To this end, we first design an F-UNet architecture that obtains high-frequency features by training complex parameters in the frequency domain of FFT-processed input images. And then, we present an S-UNet by integrating a spatial attention mechanism into the UNet architecture to enhance its capability of extracting spatial domain information from images. Finally, we propose a feature fusion module to integrate cross domain information, allowing for the extraction of richer and more comprehensive features. In this way, we can efficiently model a cross-domain embedding cost learning network at both spatial and frequency scales, enhancing its ability to resist steganalysis and resulting in more secure and robust steganography. Experimental results demonstrate that the proposed method exceeds current methods in steganalysis resistance.