Speech Coder Identification Using Chaotic Features Based on Steganalyzer Models

Abstract

In this study a steganalyzer model based on chaotic features is adapted to codec identification problem. As conventional steganalyzers detect suspected packages buried in bitstreams using statistical analysis, the proposed “speech coder identification” model also reveals the statistical differences of the outputs produced by different speech codecs. Essentially, the design of a coder/codec identifier is equivalent to a classifier training where the chaotic features extracted from the suspicious bitstream samples are used as inputs. During training, weight values which dissociate the codec types best are investigated. Finally, performances of the trained classifiers are evaluated by using test sets. According to the test results, for a closed group which is formed from nine different output types, polynomial SVM classifiers have identified more than 97% of the samples correctly.