Blind code identification using deep neural networks

The non-hierarchical communication paradigm shift in the next generation wireless networks reveals the requirement for the adaptive configuration of communication parameters in order to provide flexibility across different scenarios and changing channel conditions. Hence, automatic identification of communication parameters will play a significant role in 5G and beyond networks. In automatic modulation and coding (AMC) systems, the transmission parameters are changed dynamically in order to adapt to the changing channel conditions. These parameters are signalled through control channels which results in an increase of the resource usage. Blind decoding is fast becoming a key instrument in such systems since it eliminates the need for signaling of parameters. Blind decoding systems are composed of two stages: code identification and decoding with the identified code parameters. There is a considerable amount of research on the application of deep learning (DL) to channel decoding, and it has been shown that DL models are able to learn the code structure. In this thesis, we consider the use of DL for blind code identification, which is a first in this domain to the best of our knowledge. It is possible to design a universal code identification system via deep neural networks due to their independence of the channel code type. The presented approach is applied in detection of widely used convolutional, turbo and polar codes in order to show its capabilities. For each code type, the DL classification models provide not only higher detection rates, but also lower and predictable delay amounts compared to the existing methods. Additionally, an analysis on the required number of codewords for model training is provided.