Robust Linguistic Steganalysis Under Distribution Shift

Robust detection methods for changing domains, limited supervision, and one-class settings.

Research Question

Linguistic steganalysis aims to detect hidden information in generated text. In practice, a detector may encounter domains, generation methods, or class distributions that differ substantially from its training data. My work investigated how detection systems can remain effective under these forms of distribution shift.

Research Progression

This research developed from unsupervised domain adaptation toward increasingly constrained settings, including source-free adaptation, few-shot learning, and one-class steganalysis. Across these settings, I explored adversarial adaptation, contrastive representation learning, and clustering-driven pseudo-labeling to reduce reliance on fully matched and labeled training data.

Selected Publications

The complete publication list is available on the Publications page.