SRIS-Net: a robust image steganography algorithm based on feature score maps

Abstract

Image steganography algorithms based on deep learning are often trained using either spatial- or frequency-domain features. It is difficult for features from a single domain to comprehensively express the content of an entire image, which usually leads to poor performance because steganography is commonly multi-task. To solve this problem, this paper proposes a robust image steganography algorithm based on feature score maps, called the secure and robust image steganography network (SRIS-Net). First, instead of spatial-domain steganography, our proposed algorithm utilizes a convolutional neural network to obtain shallow spatial-domain features. These features are decomposed by Laplacian pyramid frequency-domain decomposition (LPFDD) to hide secret information in the different frequency sub-bands with a progressive assisted hiding strategy that significantly reduces the influence of the secret information on the cover image, achieving significant invisibility and robust performance. In addition, we propose a global–local embedding module (GLEM) to achieve embedding by considering the overall structure of the image and the local details, and a dual multi-scale aggregation sub-network (DMSubNet) to perform multi-scale reconstruction to improve the quality of the carrier image. For security, we propose a dual-task discriminator structure, while giving a real/fake judgment of the image, which can generate a feature score map of the cover image's region of interest (ROI) to guide the embedding module to generate a carrier image with higher imperceptibility and undetectability. Experimental results on BOSSBase show that our SRIS-Net outperforms mainstream methods in terms of undetectability and robustness, with more than 9.2 and 3.4 dB improvement in visual quality, respectively, and the capacity can be increased up to approximately 72–96 bits per pixel.

1. Introduction

Image steganography (Cheddad et al., 2010; Wengrowski and Dana, 2019; Tancik et al., 2020) hides information by exploiting image redundancy. A secret image is embedded in a cover image to generate a carrier image, while maintaining better visual quality of the carrier image and higher accuracy of the secret image extraction. Based on security considerations, the carrier image is usually required to be visually indistinguishable from the cover image. Traditional methods often hide the secret information using the least significant bit (LSB) (Barni et al., 2001; Li XL et al., 2009). These methods can hide only a limited amount of information and are not robust enough against various attacks, such as steganalysis detection (Ren et al., 2025).

In recent years, deep learning (DL)-based image steganography has shown promising results (Hu et al., 2018; ur Rehman et al., 2018). Baluja (2017) proposed the first convolutional neural network (CNN) in this domain. Researchers further improved image steganography performance by improving the network structure (Baluja, 2020) and introducing new loss functions (Singh et al., 2022). Invertible neural networks (INNs) demonstrate superior performance in image steganography (Lu et al., 2021; Yang et al., 2024) due to their precise fulfillment of the inverse relationship between embedding and extraction processes. Some studies (Chen et al., 2020) focused on designing frameworks with good robustness, and other works improved robustness by introducing noise in training (Ying et al., 2022).

Attention-based data hiding using a generative adversarial network (ADH-GAN) (Yu, 2020) guarantees both robustness and capacity. However, extensive training of an attention module is required to obtain attention weight maps for steganography assistance. Inspired by PatchGAN (Isola et al., 2017), a discriminator can evaluate the entire image generated by the generator and feed the attention back to each patch. As for the image steganography task, if the region of interest (ROI) of the steganalysis can be known before embedding, higher security can be achieved. The above methods directly concatenate or add the cover image and the secret image, or extract their features and then concatenate or add them, without considering the correlation and adaptability between features, and thus it is not optimal for the network to excel in only some aspects of performance but perform poorly in others. Moreover, image steganography based on DL is usually trained using features in a single domain, such as the spatial (Lu et al., 2021) or frequency (Jing et al.