Federated vs. Centralized Machine Learning under Privacy-elastic Users: A Comparative Analysis

The proliferation of machine learning (ML) applications has lately witnessed a considerable shift to more distributed settings, even reaching hand-held mobile devices; there, contrary to typical Centralized learning (CL) whereby the involved (large amounts of) training data are centrally gathered to train models, the load of training tasks is distributed across a set of capable mobile learners at the expense of their own energy. The idea of Federated learning (FL) has emerged as a privacy-preserving mechanism suggesting that the ML model parameters rather than data, are sent over the network to a central point of aggregation. However, when relaxing the privacy concerns, the debate strongly relates to the available network resources. Interestingly, the sofar theoretical or even experimental comparison of the two approaches overlooks network conditions and remains of low realism. In this work we rely on past measurement studies to introduce a realistic system model that accounts for all involved mobile network conditions such as bandwidth and data availability (af-fecting training accuracy and model aggregation) as well as user mobility patterns (affecting data loss). A dedicated simulation framework we have developed replays rich mobile-traces allowing for a comprehensive comparison of the two ML approaches over a large set of training data shedding light on network-resources utilization, energy efficiency and training convergence. Intuitively, our results suggest that the ratio between the employed raw data and the corresponding ML model shapes the conditions under which FL acts as a network-efficient alternative to CL. Interestingly enough, asymmetry in data availability across users as well as their varying number are shown to hardly affect the FL approach in traffic and energy needs, pointing both to its promising potential and the need for further research.