Feature-based encoder classification of compressed audio streams

Abstract

Today's digital audio coding algorithms use sophisticated models to maximise the encoding rate with minimal audible distortion. As a result of this complexity, different implementations of one encoding standard tend to produce varying output streams for the same uncompressed input data. This article presents a method to distinguish between encoding programs used to compress ISO/MPEG 1 Audio Layer-3 (MP3) files on the basis of statistical features that can be extracted from the compressed streams. The method employs a Bayesian machine learning classifier to determine the most likely encoder from a vector of 10 features. Experimental evidence suggests that the method is reliable enough to decrease the rate of false positives in a stego-detection case. Thus, it can be considered as a generic tool to increase the overall reliability of steganalysis in MP3 files. Moreover, a post-hoc interpretation of the trained classifier's parameters reveals interesting details about the degree of relation between subsets of the 20 encoding programs examined in the study. Further topics, such as implications on the robustness and possible extensions to different file formats are addressed in the discussion.