Enhancing image steganalysis via multi-source domain invariant feature learning
摘要
Image steganalysis faces significant challenges due to cover source mismatch (CSM) caused by diverse steganographic algorithms. This paper proposes a model that learns both intra-domain- and inter-domain-invariant features across multiple source domains to improve generalization. Distillation learning extracts rich spatial and frequency domain features, while second-order statistics alignment minimizes feature discrepancies. Dynamic weight assignment balances optimization objectives, enhancing model performance. Experimental results demonstrate up to a 4.87% improvement in detection accuracy under mismatched scenarios, showcasing the model’s practical applicability. The relate code and datasets will be posted at https://github.com/ailiaoye/msd.git.