Deep Learning Image Steganography Based on Dual-Path Fusion in Frequency and Spatial Domains

Contemporary deep learning-based image steganography techniques for embedding images within images are hindered by inadequate utilization of frequency-domain features and limited steganographic security, restricting their effectiveness in practical privacy protection contexts. To mitigate these limitations, we introduce a frequency–spatial dual-path fusion-based deep steganography approach, termed FS-Stego. This method incorporates a frequency–spatial dual-path architecture within the generator network. Specifically, the frequency-domain processing module facilitates feature embedding in the complex domain, while the spatial-domain processing module maintains the image’s structural integrity, thereby enabling the co-optimization of multi-dimensional features. Second, an adaptive fusion module is developed to dynamically adjust the weights of the two paths, while residual connections and attention mechanisms are utilized to mitigate feature loss. Third, a multi-objective loss function is implemented to simultaneously optimize the quality of the stego images and the reconstruction accuracy of the secret images. The proposed method utilizes three open-source datasets as cover images and the LFW dataset as the secret images. Experimental results demonstrate that, compared to existing deep steganographic techniques, the stego and recovered images achieve superior peak signal-to-noise ratios (PSNR) and structural similarity (SSIM). Regarding model efficiency, the number of parameters is reduced to below 0.98 million, significantly enhancing practical performance. The proposed method ensures high-quality image recovery while maintaining steganographic security, thereby offering an effective solution for privacy protection.