Steganalysis Using Partially Ordered Markov Models

摘要

The field of steganalysis has blossomed prolifically in the past few years, providing the community with a number of very good blind ste-ganalyzers. Features for blind steganalysis are generated in many different ways, typically using statistical measures. This paper presents a new image modeling technique for steganalysis that uses as features the conditional probabilities described by a stochastic model called a partially ordered Markov model (POMM). The POMM allows concise modeling of pixel dependencies among quantized discrete cosine transform coefficients. We develop a steganalyzer based on support vector machines that distinguishes between cover and stego JPEG images using 98 POMM features. We show that the proposed steganalyzer outperforms two comparative Markov-based steganalyzers [25,6] and outperforms a third steganalyzer [23] on half of the tested classes, by testing our approach with many different image databases on five embedding algorithms, with a total of 20,000 images.