Abstract
In this paper, we study the wavelet decomposition based steganalysis technique due to Lyu and Farid. Specifically we focus on its performance with JSteg steganograpy. It has been claimed that the Lyu-Farid technique can defeat JSteg; we confirm this using different images for the training and test sets of the SVM classifier. We also show that the technique heavily depends on the characteristics of training and test set. This is a problem for real-world implementation since the image source cannot necessarily be determined. With a wide range of image sources, training the classifier becomes problematic. By focusing only on different camera makes we show that steganalysis performances significantly less effective for cover images from certain sources.