Feature Reduction Algorithm for Universal Steganalysis

Abstract

Steganography is the practice of communicating a secret message by hiding it in a cover medium like an image file. Steganalysis is the detection of steganography. There are two kinds of steganalysis approaches: specific steganalysis where the steganalyst has knowledge about the algorithm of steganography used to hide data and universal (blind) steganalysis where he has no idea about the algorithm used to hide data. In this paper, we use an image as support for steganalysis. In the case of blind steganalysis, a lot of features are extracted from the image to train a classifier. The great number of features makes computation very difficult and, by the way, prevents the implementation of certain supervised learning algorithms to build models for universal steganalysis. In this work, we propose a universal steganalysis method based on a new way of selecting the most relevant features to train a classifier. This algorithm consists of choosing the best set of features, generated by PFA (Principal Feature Analysis) for unsupervised learning [6], which makes it possible to obtain the best result for k-means with two classes \( (k \leftarrow 2) \). After finding the best set of features we use it to train a model for binary classification (cover vs stego). This approach gave good results in our experiments and open a new way of doing universal steganalysis without expensive computation resources.