Deep Learning Enabled Methods for Information Hiding

Abstract

When you look at the provided image, what do you see? Two gorgeous fur babies? Funny looking life jackets? To the unassuming eye, this image is exactly as it appears---two gorgeous dogs in life jackets---but beneath the surface, this image contains the entire work of Shakespeare's "Hamlet", embedded in a manner to be imperceptible to the human eye. Information hiding is a method of secure communication through strategic concealment and extraction. It enables communication while completely disguising the act itself. Because information hiding techniques are deeply intertwined with evolving technologies, they advance as rapidly as new innovations emerge. As methods for embedding hidden data become more sophisticated, so too do the techniques for detecting and eliminating them, creating an ongoing cycle of adaptation and countermeasures.

This dissertation examines information hiding through the lens of deep learning, exploring topics such as deep learning-based sanitization techniques for removing hidden information from media, automated methods for visible watermark removal, trigger-based fragile watermarking schemes for safeguarding model integrity, and an analysis of the robustness of modern malware detection models.