An Iterative Learning Algorithm for Deciphering Stegoscripts: a Grammatical Approach for Motif Discovery

Steganography, or information hiding, is to conceal the existence of messages so as to protect their confidentiality. We consider deciphering a stegoscript, a text with secret messages embedded within a covertext, and identifying the vocabularies used in the messages, with no knowledge of the vocabularies and grammar in which the script was written. Our research was motivated by the problem of identifying conserved non-coding functional elements (motifs) in regulatory regions of genome sequences, which we view as stegoscripts constructed by nature with a statistical model consisting of a dictionary and a grammar. We develop an iterative learning algorithm, WordSpy, to learn such a model from a stegoscript. The model then can be applied to identify the embedded secret messages, i.e., the functional motifs. Our algorithm can successfully recover the most possible text of the first ten chapters of a novel embedded in a stegoscript and identify the transcription factor binding motifs in the upstream regions of ∼ 800 yeast genes.