HTLIN-Stega: Hierarchical Text-Label Integration Network for Text Steganalysis

Neural text steganalysis has achieved impressive success. However, existing algorithms struggle to detect mixed texts generated by multiple steganography algorithms. To improve the detection capability for mixed texts, this paper proposes a novel Hierarchical Text-Label Integration Network (HTLIN-Stega), introducing the steganography-related hierarchical structure into the training process. Specifically, we assign different granularities of hierarchical labels to text samples, ranging from coarse-grained to fine-grained labels. Then a Word-Label Fusion Unit is proposed to map the text and multi-layer label features into a joint space. A loss function is designed to align labels and text features within the joint space and to learn the distance relationships between text representations and labels from different granularities, including coarse-grained labels, fine-grained labels, sibling labels, and inter-class labels. Experimental results on various hierarchical datasets show that the HTLIN-Stega can effectively learn multi-layer steganographic features and significantly improve the detection accuracy of mixed steganographic text.