Linguistic Steganalysis Method Based on Meta-learning and Domain Adaptation
Generative linguistic steganography brings challenges to linguistic steganalysis while generating steganographic text with high naturalness.The neural-based methods enable effective detection when the source domain(training text)and the target domain(tes-ting text)belong to the same training corpus and the same steganography algorithm.However,we cannot predict which training corpus and steganography algorithm are used to generate the text to be detected in practical applications,thus most of these neural-based methods are difficult to be practical.To address this issue,we introduce the basic ideas of meta-learning and domain adaptation to improve the generalization of the steganalysis model.Moreover,we construct a word importance semantic encoding module with the pre-trained language representation model RoBERTa to fully extract text semantic features.Given these features,we further propose a multi-scale perception module and an attention mechanism to capture the word correlation changes that exist between adjacent words and nonadjacent words caused by steganography.The experimental results demonstrate that in multiple cross-domain scenarios,this method has shown an average improvement of 9％in detection accuracy compared to the existing text steganalysis method Fs-Stega.
generative linguistic steganalysiscross-domain detectiondomain adaptationfew-shot learningmeta-learning
万方数据
维普
