Post-Training Representation Editing for Cross-Domain Linguistic Steganalysis

A novel approach known as REED (Representation Editing) has been introduced for cross-domain linguistic steganalysis. In practical applications, the texts analyzed frequently originate from unfamiliar domains that exhibit varying vocabularies, topics, writing styles, and steganographic patterns, which hampers detection efficacy. Current cross-domain techniques rely on distribution alignment and domain-invariant feature learning but often fall short. REED functions after training: the detector is initially trained on source-domain data, followed by freezing the feature extractor and classifier, then deterministically adjusting intermediate representations prior to classification. For domain adaptation, a domain-offset vector is created from marginal representations of both source and target. A source-domain vector is utilized for domain generalization. This research is available on arXiv under ID 2605.28298.