Packet-level prediction of mobile-app traffic using multitask Deep Learning

The prediction of network traffic characteristics helps in understanding this complex phenomenon and enables a number of practical applications, ranging from network planning and provisioning to management, with security implications as well. A significant corpus of work has so far focused on aggregated behavior, e.g., considering traffic volumes observed over a given time interval. Very limited attempts can instead be found tackling prediction at packet-level granularity. This much harder problem (whose solution extends trivially to the aggregated prediction) allows a finer-grained knowledge and wider possibilities of exploitation. The recent investigation and success of sophisticated Deep Learning algorithms is now providing mature tools to face this challenging but promising goal. In this work, we investigate and specialize a set of architectures selected among Convolutional, Recurrent, and Composite Neural Networks, to predict mobile-app traffic at the finest (packet-level) granularity. We discuss and experimentally evaluate the prediction effectiveness of the provided approaches also assessing the benefits of a number of design choices such as memory size or multi-modality, investigating performance trends at packet level focusing on the head and the tail of biflows. We compare the results with both Markovian and classic Machine Learning approaches, showing increased performance with respect to state-of-the-art predictors (high-order Markov chains and Random Forest Regressor). For the sake of reproducibility and relevance to modern traffic, all evaluations are conducted leveraging two real human-generated mobile traffic datasets including different categories of mobile apps. The experimental results witness remarkable variability in prediction performance among different apps categories. The work also provides valuable analysis results and tools to compare different predictors and strike the best balance among the performance measures.