Multiclass Neural Network for Codec Classification

Abstract

In this paper, we suggest to remove or modify the denoted class of codec in a bitstream for military purposes in image communication. In that case, a decoder first needs to determine the codec type to restore the original data. This paper proposes a codec classification method which has not been studied much yet. For extracting the feature of a bitstream, Recurrent Neural Network (RNN) model is used since it is suitable for time series data used for training on classification. Video codecs have their own distinctive header structures, which can be considered features in the encoded bitstreams. The proposed method extracts the feature of an encoded bitstream and classifies the bitstream into the specific codec. Three standard codecs, MPEG-2, H.263, and H.264/AVC, are used to generate the training and the test data set in the experiment. We analyze several components affecting the performance and compare to conventional algorithm. The performance degrades when two kinds of bitstreams generated by H.263 and H.264/AVC are trained together. However, when the training data includes both H.263 and H.264/AVC, performances improved with increasing training data set sizes.