Research On Quantitative Evaluation Of Steganalytic Algorithms And Application

As information hiding technology has been developing rapidly, more and more digital steganographic software come forth in the Internet. Steganographic software can be easily used in activities which disserve society and national security by lawbreakers or terroristic organizations. Therefore detecting technology of hidden information, namely the first step of steganalysis becomes a hotspot in the information security. So far, most work focus on the design of steganalytic algorithms. While few work refer to the evaluation of steganalytic algorithms in steganalysis. Nevertheless, reasonable evaluation of steganalytic algorithms is helpful in improving algorithms, optimizing and combining a great deal of algorithms to design steganalysis system conforming practical application. The main work of this paper is to study quantitative evaluation of steganalytic algorithms and interrelated application.Firstly, we present a scenario of quantitative evaluation combining ROC analysis and according to characteristic of steganalytic algorithms as well as the need of application. The scenario contains four evaluation rules relating to reliability and veracity, applicability, classifying cost and computing complexity of steganalytic algorithms. The scenario discusses minimum embedded ratios of stego objects which algorithms can detect efficiently. At the same time, we analyze the factor which influences computing complexity of algorithms. Accordingly, we advance analysis of sample size. Aimed for different requirement in FAR (False Alarm Rate) and FDR (False Detection Rate), we advance the analysis of classifying cost.Secondly, we apply the scenario of quantitative evaluation to six LSB (Least Significant Bit) steganalytic algorithms. We set up 8 different types of BMP grayscale images (33272 images totally) so that the gained results are credible. We make a great deal of experiment to get evaluation results of six algorithms. Furthermore, the results are compared from different sides so that the best algorithm can be determined.In the last part, we bring forward two algorithms methods based on many steganalytic algorithms. Respectively, two methods that we advance are based on majority theory and bayes classification model. And their theoretical frameworks are also given. Theoretical error analysis indicates: the former method's error reduce exponentially as the number of steganalytic algorithms increases, and the later method's error is related to the number of steganalytic algorithms and the error of single steganalytic algorithm. They can both improve detection effect from different degree. Furthermore, comparison of two methods as well as their different applying situation is given.