Abstract:

The performance of joint photographic experts group (JPEG) image steganalysis is related not only to the information-hiding methods but also to the complexity of image content. A novel approach based on pre-classification and feature extension is proposed to improve the accuracy of JPEG universal steganalysis. Firstly, an image content complexity evaluation criterion is exploited to divide the training samples into several overlapping categories. After that, the more effective extended feature set is extracted from each category respectively to build a classifier. Given a testing image, its category is determined according to the minimum distance from its complexity to the average value of each category. Then, the feature set is extracted and sent to the corresponding classifier. Experimental results indicate that this new approach exhibits excellent performance for the detection of typical JPEG steganography algorithms.