Quantitative Steganalysis Of Digital Images Using Ensemble Learning

Quantitative steganalysis of digital images is a hotspot and difficult task in the research of steganalysis. Currently, many outstanding works on the quantitative steganalysis of digital images have been achieved, but there are still many problems need to be solved, such as the “Curse of Dimensionality”, the difficulty in capturing the non-linear relationship between the embedding relative payloads and the statistical features, the lack of adaption of the statistical feature to the steganography algorithms. This thesis carries out the study on the problems mentioned above, and the contents are as follows:1. Targeting at “Curse of Dimensionality” problem, the quantitative steganalysis framework based on ensemble learning was presented. Firstly, multiple training sets were generated using random subspace and bootstrap method and multiple base estimators were trained on them. Then, an diversity regularized ensemble pruning algorithm was used to find a balance between the size of the ensemble and the accuracy of it. Finally, the final estimation result is the weighted combination of the results from the multiple base estimators.2. In order to capture the non-linear relationship between the embedding relative payloads and the statistical features, a construction method of the estimator ensemble based on support vector regression was proposed. Under the quantitative steganalysis framework based on ensemble learning, the proposed detailed construction method consist of the setting of the dimensionality of the random subspace, the number of the base estimator, the parameter for the ensemble pruning. In the experiment, the estimator ensembles for three embedding algorithms, namely EA, HUGO, WOW, were conducted. Compared to existed typical methods, the estimator ensemble constructed by our proposed method achieved higher estimation accuracy.3. Aiming to enlarge the sensitivity of the features to the embedding, a quantitative steganalysis algorithm for adaptive steganography based on selection of blocks was proposed. The point of this algorithm is estimating the areas with high change rate and extraction of the statistical feature from these areas. Meanwhile, we theoretically prove that the change rate of the areas we estimated is higher than that of the whole image. Then, we experimentally verified that the features extracted from these areas are more sensitive to the embedding. In the experiment, quantitative steganalysis of S-UNIWARD and J-UNIWARD is carried out. The experimental results indicate that, using the features extracted from the estimated areas could achieve better quantitative steganalysis performance than using the features extracted from the whole image.Finally, a general conclusion of the contribution of this thesis is made and some other research topics that need to be further studied are presented.