@ARTICLE{Ashari, author = {Ashari, Fatemeh and Riahi, Nooshin and }, title = {Audio Steganalysis based on Inter-frame correlation and recursive feature elimination}, volume = {15}, number = {3}, abstract ={Dramatic changes in digital communication and exchange of image, audio, video and text files result in a suitable field for interpersonal transfers of hidden information. Therefore, nowadays, preserving channel security and intellectual property and access to hidden information make new fields of researches naming steganography, watermarking and steganalysis. Steganalysis as a binary classification distinguish clean signals from stego signals. Features extracted from time and transform domain are proper for this classifier. Some of steganalysis methods are depended on a specific steganography algorithm and others are independent. The second group of methods are called Universal steganalysis. Universal steganalysis methods are widely used in applications because of their independency to steganography algorithms. These algorithms are based on characteristics such as distortion measurements, higher order statistics and other similar features. In this research we try to achieve more reliable and accurate results using analytical review of features, choose more effective of them and optimize SVM performance. In new researches Mel Frequency Cepstral Coefficient and Markov transition probability matrix coefficients are used to steganalysis design. In this paper we consider two facts. First, MFCC extract signal features in transform domain similar to human hearing model, which is more sensitive to low frequency signals. As a result, in this method there is more hidden information mostly in higher frequency audio signals. Therefore, it is suggested to use reversed MFCC. Second, there is an interframe correlation in audio signals which is useful as an information hiding effect. For the first time, in this research, this features is used in steganalysis field. To have more accurate and stable results, we use recursive feature elimination with correlation bias reduction for SVM. To implement suggested algorithm, we use two different data sets from TIMIT and GRID. For each data sets,Steghide and LSB-Matching steganography methods implement with 20 and 50 percent capacity. In addition, one of the LIBSVM 3.2 toolboxes is sued for implementation. Finally, the results show accuracy of steganalysis, four to six percent increase in comparison with previous methods. The ROC of methods clearly shows this improvement. }, URL = {http://jsdp.rcisp.ac.ir/article-1-625-en.html}, eprint = {http://jsdp.rcisp.ac.ir/article-1-625-en.pdf}, journal = {Signal and Data Processing}, doi = {10.29252/jsdp.15.3.113}, year = {2018} }