Least significant bit steganography detection with machine learning techniques

Steganography is the art to transmit secret messages through seemingly innocuous files, while steganalysis is to detect or extract hidden messages from such files. Heuristic and feature based methods were proposed for hidden information detection, but their methods seem to be too specific. In this paper we propose a general framework of applying machine learning to steganalysis for LSB (Least Significant Bit steganography) hidden information detection. We have investigated the performance of our method on different classification methods, different complexities of images and different embed rates. Feature based classifiers are trained for sequential and non-sequential LSB hidden information detection. The results show that chi-square feature based decision tree can get almost 10 percentage higher accuracy than the simple threshold based chi-square techniques for sequential LSB detection. In the non-sequential case, just 2-D RS steganalysis features based decision trees can achieve an accuracy of 97.56% in mixed embed rate case.