Combating highly imbalanced steganalysis with small training samples using feature selection,

We consider a particular paradigm of steganalysis, namely, highly imbalanced steganalysis with small training samples, in which the cover images always significantly outnumber the stego ones. Researchers have rigorously studied sampling and learning algorithms as well as feature selection approaches to the class imbalance problem, but the research in the steganalysis domain is rare. This study provides a systematic comparison of eight feature selection metrics and of three types of methods developed for the imbalanced data classification problem in the steganalysis domain. Each metric is compared across three different classifiers and four steganalytic features. The efficiency of the metrics is evaluated to determine which performs best with minimal features selected. The performance of the three types of methods and their combinations is examined. Moreover, we also investigate the effect of feature dimensionality, sample number and imbalance degree on the performance of feature selection inresolving imbalanced image steganalysis.