Semi-supervised QIM steganalysis with ladder networks

Abstract

Recently, deep learning-based Quantization Index Modulation (QIM) steganalysis algorithms have achieved great success. However, most of them are supervised learning algorithms that rely on a large number of labeled samples and have poor generalization performance. Towards addressing the challenge, we present a novel semi-supervised ladder network, termed SSLadNet, for weak signal detection in QIM steganalysis of VoIP streams. In particular, we integrate supervised learning and unsupervised learning into an end-to-end learning architecture via a ladder network, and achieve joint optimization for semi-supervised learning by backpropagation to minimize the sum of supervised and unsupervised cost functions. To the best of our knowledge, this is the first deep learning-based semi-supervised detection model applied to QIM steganalysis that can effectively extract rich features reflecting the correlation changes between codewords caused by QIM steganography. Experimental results showed that even for the labeled samples with a number of 512, SSLadNet can achieve a detection accuracy of around 96.09% for $1000\mathrm{ms}$ long samples and 100% embedding rate, and outperforms the state-of-the-art methods based on semi-supervised learning.