In recent years, digital watermarking has emerged as an increasingly active research area. Information can be hidden into images, videos, and audios imperceptibly to human beings. It provides vast opportunities for covert communications. Consequently, methods to detect covert communications are called for. This task is especially urgent for law enforcement to deter the distribution of children pornography images/videos hidden inside normal images/videos, and for intelligence agencies to intercept communications of enemies. Steganalysis is the art and science to detect whether a given medium has hidden message in it. On the other hand, steganalysis can serve as an effective way to judge the security performance of steganographic techniques. In other words, a good steganographic method should be imperceptible not only to human vision systems, but also to computer analysis.

The huge diversity of natural images and the wide variation of data embedding algorithms make steganalysis a tough mission. However, an original cover medium and its stego-version (with hidden message inside) always differ from each other in some aspects since the cover medium is modified during the data embedding. A method designed to blindly detect stego-images is referred to as a general steganalysis method, meaning it does now know which specific data hiding method is actually used and it does not have the original image in detection. From this point of view, the general steganalysis methods have more real value for deterring covert communications.

Steganalysis as a Task of Pattern Recognition

Based on whether an image contains hidden message, images can be classified into two classes: the image with no hidden message and the corresponding stego-image (the very image but with message hidden in it). Steganalysis can thus be considered as a pattern recognition process to decide which class a test image belongs to. The key issue for steganalysis just like for pattern recognition is feature extraction. The features should be sensitive to the data hiding process. In other words, the features should be rather different for the image without hidden message and for the stego-image. The larger the difference, the better the features are. The features should be as general as possible, i.e., they are effective to all different types of images and different data hiding schemes. Often in practice it is very hard to achieve a high recognition rate with a single feature when the classification such as steganalysis is complicated in nature. Therefore, M-D feature vectors should be used under the circumstances. Each image is a sample point in the M-D feature space. Steganalysis has thus become a pattern classification in the M-D feature space. It is desirable to have features in individual dimensions of the feature vector independent or at least less related to one another. Just like for pattern recognition, in addition to feature extraction, classifier design is another key issue for steganalysis; and the performance of a steganalysis system, combination of feature extraction and classifier design, is evaluated by its classification success or error rate.  

กก