Multiple linear regression for universal steganalysis of images

Abstract

Steganography is the art of hiding information in a cover (carrier) medium to obtain a stego-medium without any suspicion from a viewer who see that last one. Steganalysis is the opposite discipline. Its goal is to detect the presence of hidden information from a stego-medium. The medium can be an audio, video or image file. In this work, we focus on image file medium. Universal steganalysis is the detection of hidden data without knowing the algorithm used to embed the message inside the carrier. There are some methods of classification between stego and cover medium proposed in literature. In this paper, we propose a new universal steganalysis method based on unsupervised and supervised machine learning algorithms. Our method reduces the cover-source mismatch problem in the first stage and uses multiple linear regression in the second stage to predict the relative payload (in terms of bits per non-zero AC DCT coefficient) of the embedded message. With this measure, we can easily calculate the length of the embedded message. In our experiments, we got reliable models in all the clusters to predict the relative payload for cover-images and stego-images.