Proposal / Submission Type

Peer Reviewed Paper

Location

Richmond, Virginia

Start Date

30-5-2012 10:30 AM

Abstract

The detection of hidden messages in JPEG images is a growing concern. Current detection of JPEG stego images must include detection of double compression: a JPEG image is double compressed if it has been compressed with one quality factor, uncompressed, and then re-compressed with a different quality factor. When detection of double compression is not included, erroneous detection rates are very high. The main contribution of this paper is to present an efficient double-compression detection algorithm that has relatively lower dimensionality of features and relatively lower computational time for the detection part, than current comparative classifiers. We use a model-based approach for creating features, using a subclass of Markov random fields called partially ordered Markov models (POMMs) to modeling the phenomenon of the bit changes that occur in an image after an application of steganography. We model as noise the embedding process, and create features to capture this noise characteristic. We show that the nonparametric conditional probabilities that are modeled using a POMM can work very well to distinguish between an image that has been double compressed and one that has not, with lower overall computational cost. After double compression detection, we analyze histogram patterns that identify the primary quality compression factor to classify the image as stego or cover. The latter is an analytic approach that requires no classifier training. We compare our results with another state-of-the-art double compression detector.

Keywords: steganalysis; steganography; JPEG; double compression; digital image forensics.

 
May 30th, 10:30 AM

Double-Compressed JPEG Detection in a Steganalysis System

Richmond, Virginia

The detection of hidden messages in JPEG images is a growing concern. Current detection of JPEG stego images must include detection of double compression: a JPEG image is double compressed if it has been compressed with one quality factor, uncompressed, and then re-compressed with a different quality factor. When detection of double compression is not included, erroneous detection rates are very high. The main contribution of this paper is to present an efficient double-compression detection algorithm that has relatively lower dimensionality of features and relatively lower computational time for the detection part, than current comparative classifiers. We use a model-based approach for creating features, using a subclass of Markov random fields called partially ordered Markov models (POMMs) to modeling the phenomenon of the bit changes that occur in an image after an application of steganography. We model as noise the embedding process, and create features to capture this noise characteristic. We show that the nonparametric conditional probabilities that are modeled using a POMM can work very well to distinguish between an image that has been double compressed and one that has not, with lower overall computational cost. After double compression detection, we analyze histogram patterns that identify the primary quality compression factor to classify the image as stego or cover. The latter is an analytic approach that requires no classifier training. We compare our results with another state-of-the-art double compression detector.

Keywords: steganalysis; steganography; JPEG; double compression; digital image forensics.