Know your Big Data Trade-offs when Classifying Encrypted Mobile Traffic with Deep Learning

The spread of handheld devices has led to the unprecedented growth of traffic volumes traversing both local networks and the Internet, appointing mobile traffic classification as a key tool for gathering highly-valuable profiling information, other than traffic engineering and service management. However, the nature of mobile traffic severely challenges state-of-art Machine-Learning (ML) approaches, since the quickly evolving and expanding set of apps generating traffic hinders ML-based approaches, that require domain-expert design. Deep Learning (DL) represents a promising solution to this issue, but results in higher completion times, in turn suggesting the application of the Big-Data (BD) paradigm. In this paper, we investigate for the first time BD-enabled classification of encrypted mobile traffic using DL from a general standpoint, (a) defining general design guidelines, (b) leveraging a public-cloud platform, and (c) resorting to a realistic experimental setup. We found that, while BD represents a transparent accelerator for some tasks, this is not the case for the training phase of DL architectures for traffic classification, requiring a specific BD-informed design. The experimental setup is built upon a three-dimensional investigation path in the BD adoption, namely: (i) completion time, (ii) deployment costs, and (iii) classification performance, highlighting relevant non-trivial trade-offs.