GISNet: Lightweight image-to-image steganography based on improved emd and dual-domain graph convolutional network

Image Steganography Technology embeds secret images into cover images to protect privacy without raising suspicion from third parties. However, methods limited to single-domain processing struggle to simultaneously achieve concealment, security, and efficiency, while high-performance models are often too complex for lightweight deployment. To address these issues, this paper proposes a lightweight image steganography network model named GISNet, based on improved Empirical Mode Decomposition and dual-domain graph convolution. The model employs an enhanced Empirical Mode Decomposition module to adaptively decompose the secret image into multiple scales and utilizes multi-scale spatial-frequency blocks along with graph convolution modules to achieve deep integration and collaborative optimization of spatial and frequency domain features. The experimental results demonstrate that GISNet outperforms existing mainstream steganographic models. On the DIV2K dataset, the Peak Signal-to-Noise Ratio (PSNR) of cover/steganographic images and secret/recovered images increases by 1.77 dB and 4.65 dB, respectively. Moreover, GISNet exhibits notable lightweight characteristics, with only 8.00M parameters, 7.88 GFLOPs of computational cost, and an inference time of 76 ms, which demonstrates its high efficiency and suitability for resource-constrained environments.