Semi-supervised Learning Based JPEG Steganalysis

The objective of steganalysis is to detect secret messages hidden in digital objects by analyzing the microscopic statistical characteristics of typical Web images. So we can use steganalysis to defeat terrorist purposes from the potential use for legitimate purposes. Most of the traditional blind steganalysis extract a conglomerate of features that can potentially detect arbitrary steganographic method, and trained a classifier or classifiers to detect each of test images. Most of the blind steganalysis systems’performances heavily rely on large-scale labeled training data, which is expensive and sometimes impossible to acquire. When the number of labeled images is insufficient, how to use a great number of unlabeled images in the field of network and social life to improve the accuracy of steganalysis systems has become an urgent problem. In this paper, the problem of how to use very few labeled data and fully exploit a large number of unlabeled data is addressed by a novel image representation named ensemble projection based JPEG steganalysis. Furthermore, improve it use co-training. All these research results will be a strong support to design new steganalysis techniques with higher detecting rates.Unlike previous methods to regularize classifying boundaries with unlabeled data, our method learns a new image representation from all available data (labeled and unlabeled) and performs plain supervised learning with the new feature. First take full advantage of all labeled data to construct detectors and detect all of the available data. An ensemble of image prototype set is selected from the test result, to represent a rich set of category information. Discriminative functions are then learned on these prototype sets. Finally, images are represented by the concatenation of their projected values onto the prototypes (similarities to them) for further classification. We steganalyze five modern steganographic methods and compare our method with classical MFS-274and Rich Model which with good analytical performance. Experiments show that ensemble projection based JPEG steganalysis can effectively exploit unlabeled images to achieve better performance, our method consistently outperforms previous methods for JPEG steganalysis when the number of labeled data is less than50. Especially for J-UNIWARD scheme, when labeled samples are50, MFS-274does not have any judgment ability, Rich Model accuracy is51.89%, accuracy of our method is74.44%.The extracted features of Ensemble projection based JPEG steganalysis can be applied to other machine learning method. We combined it with co-training. Firstly, because the extracted features are essentially probability value, features are independent of each other, so can be divided into two view, and use labeled data initialized two classifiers on two views. Secondly, iteratively trains two classifiers on the two views. In every iterative round, based on specific data editing techniques to explicitly estimate confidence of either classifier’s predictions and uses they to augment the training set of the other. Thirdly, use specific threshold value to avoid introducing undesirable classification noise. Finally, stop the co-training process automatically when either classifier’s predictive error on original labeled set increase, or the expected predictive errors of both classifiers won’t decrease. Experiments on several data sets across four steganography method show that co-training can effectively exploit unlabeled images to achieve better performance when annotations of images is scarce. Even at10labeled data,990test data, the improvement of our method compared to MFS-274the correct accuracy can be up to5.96%on MB1steganography.