A Lightweight Multi-Domain Fusion Approach for Image Steganalysis and Stegomalware Detection
Abstract
The rapid evolution of image steganography has introduced increasingly sophisticated techniques for concealing covert payloads within seemingly benign digital media, creating significant challenges for digital forensics and cybersecurity. Indeed, such hidden information is extremely hard to detect because current embedding techniques create only slight statistical changes in images while maintaining their visual integrity. Moreover, modern deep learning networks usually experience semantic bias due to their excessive reliance on visual semantics which is more important than forensic features for accurate steganalysis. In this paper, a multi-domain feature fusion model is developed that incorporates spatial domain-based residual analysis, wavelet transforms in the frequency domain, and deep semantic features into the model in order to enhance detection of hidden information in JPEG images. The model utilizes SRM residual filters, DWT branch, and EfficientNet-B0 with attention mechanism to analyze complementary forensics in different representation modalities. Gated fusion is used to fuse features of all three streams into the final representation before classification. The model was tested on the ALASKA2 dataset consisting of cover and stego images created by state-of-the-art embedding techniques.