Visual Detection of High Capacity Natural Image Steganography: Is A Known Cover Effective?

In terrorist and criminal investigations, law enforcement agencies may be called upon to visually inspect images that may or may not contain embedded steganographic data. We compared the accuracy of two different approaches to visual steganalysis using signal detection theory: a same/different task, where comparisons were made to a known cover, and a yes/no task, where no cover was provided. Using least significant bit steganography (encoded across multiple bit layers to increase capacity) in 15 colour natural images from the VisTex database, using a controlled same/different task (N=58) and a yes/no task (N=53), the presence of a known cover resulted in generally lower accuracy, contrary to expectation: d´>1 was only observed for colour layers 5-8 in the same/different task, while d´1> was observed for layers 4-8 in the yes/no task. In summary, steganographic in the lower layers appears to be highly resistant to visual steganalysis with or without a known covers.