This paper proposes an improved steganalytic method when cover selection is used in steganography. We observed that the covers selected by existing cover selection methods normally have different characteristics from normal ones, and propose a steganalytic method to capture such differences. As a result, the detection accuracy of steganalysis is increased. In our method, we consider a number of images collected from one or more target (suspected but not known) users, and use an unsupervised learning algorithm such as k-means to adapt the performance of a pre-trained classifier towards the cover selection operation of the target user(s). The adaptation is done via pseudo-labels from the suspected images themselves, thus allowing the re-trained classifier more aligned with the cover selection operation of the target user(s). We give experimental results to show that our method can indeed help increase the detection accuracy, especially when the percentage of stego images is between 0.3 and 0.7.

Introduction

Steganography is the art of covert communication, aiming to transmit data secretly through public channels without drawing suspicion, and steganalysis aims to disclose the secret transmission by analyzing suspected media [1]. Modern steganalytic methods use supervised machine learning to investigate the models of the covers and the stegos. Features are extracted from a set of images to train a common steganalytic classifier, which is then used to distinguish real stego images from normal (cover) images without any hidden information [2], [3]. The ensemble classifier [4] is widely used to enhance the performance by using multiple classifiers. The feature extraction and machine learning based steganalysis has been proved to be efficient. The most popular feature set is SRM (Spatial Rich Model) [5], which are the fourth order co-occurrence matrices for describing the dependencies among different pixels. After SRM, some improved feature extraction methods are proposed [6], [7]. In PSRM (Projections of Spatial Rich Model) [6], neighboring residual samples are projected onto a set of random vectors and the histograms of the projections are taken as the feature. The feature set maxSRMd2 [7] is a variant of SRM that makes use of the modification probabilities of cover elements during data embedding, which is called probabilistic selection channel. Recently, deep learning based steganalysis has also achieved good performances with enough training data [8]–[۱۰].