Machine Learning in Image Steganalysis

Chapter 3

Getting Started with a Classifier

In Part II of the book, we will survey a wide range of feature vectors proposed for steganalysis. An in-depth study of machine learning techniques and classifiers will come in Part III. However, before we enter Part II, we need a simple framework to test the feature vectors. Therefore, this chapter starts with a quick overview of the classification problem and continues with a hands-on tutorial.

3.1 Classification

A classifier is any function or algorithm mapping objects to classes. Objects are drawn from a population which is divided into disjoint classes, identified by class labels. For example, the objects can be images, and the population of all possible images is divided into a class of steganograms and a class of clean images. The steganalytic classifier, or steganalyser, could then take images as input, and output either ‘stego’ or ‘clean’.

A class represents some property of interest in the object. Every object is a member of one (and only one) class, which we call its true class, and it is represented by the true label. An ideal classifier would return the true class for any object input. However, it is not always possible to determine the true class by observing the object, and most of the time we have to settle for something less than ideal. The classifier output is often called the predicted class (or predicted label) of the object. If the predicted class matches the true class, we have correct classification. Otherwise, we ...