Machine Learning in Image Steganalysis

Chapter 12

Other Classification Algorithms

The problem of classification can be approached in many different ways. The support vector machine essentially solves a geometric problem of separating points in feature space. Probability never entered into the design or training of SVM. Probabilities are only addressed when we evaluate the classifier by testing it on new, unseen objects.

Many of the early methods from statistical steganalysis are designed as hypothesis tests and follow the framework of sampling theory within statistics. A null hypothesis H₀ is formed in such a way that we have a known probability distribution for image data assuming that H₀ is true. For instance, ‘image I is a clean image’ would be a possible null hypothesis. This gives precise control of the false positive rate. The principle of sampling theory is that the experiment is defined before looking at the observations, and we control the risk of falsely rejecting H₀ (false positives). The drawback of this approach is that this p-value is completely unrelated to the probability that H₀ is true, and thus the test does not give us any quantitative information relating to the problem.

The other main school of statistics, along side sampling theory, is known as Bayesian statistics or Bayesian inference. In terms of classification, Bayesian inference will aim to determine the probability of a given image I belonging to a given class C a posteriori, that is after considering the observed image I.

It is debatable ...