Steganalysis with mismatched covers: do simple classifiers help?
Proceedings of the on Multimedia and security
Moving steganography and steganalysis from the laboratory into the real world
Proceedings of the first ACM workshop on Information hiding and multimedia security
| Hi-index | 0.00 |
Steganography is the art of covert communication. Its goal is achieved by hiding secret messages into innocuous objects such as digital images, audio files, etc. As any other technology, steganography can be used for malicious purpose. The need to detect steganographic objects give rise to steganalysis, a complementary task to steganography. The main focus of this dissertation is on steganalysis of JPEG images, as the JPEG image format is nowadays the most frequently used image format. In the first part of the dissertation, a detector recognizing six popular steganographic algorithms is presented. This detector is novel in many ways. First, it not only detects the presence of a secret message, but also assigns the image to a class according to the algorithm used to embed the message. Second, it detects stego content in single-compressed as well as in double-compressed JPEG images. Third, it offers superior accuracy in comparison to prior art. The detector is based on feature extraction and supervised training of two banks of multi-classifiers implemented by Support Vector Machines. The first bank targeted to single-compressed images contains a separate multi-classifier trained for each JPEG quality factor from a certain range. Another bank of multi-classifiers is trained for double-compressed images for the same range of primary quality factors. Multi-classifier banks are preceded by a pre-classifier detecting double-compression and estimating the primary quantization table. The second part of the dissertation presents a method for benchmarking security of steganographic algorithms. We argue that a good benchmark should be dependent only on the model chosen to represent cover and stego objects (feature set). While the KL divergence would be a preferable measure, because it is a fundamental quantity, there are practical difficulties in computing it. Therefore a Maximum Mean Discrepancy (MMD) is proposed as a measure of steganographic security, because it is well understood theoretically, and is numerically stable even in high-dimensional spaces.