Smartphones are ubiquitous in our daily lives. They form an easy-to-reach computing re- source with a direct access to a considerable amount of personal information. They represent a highly valuable source of data for telecom operators, but their highly decentralized nature and the evident customer's expectations regarding privacy require new statistical learning ap- proaches. Ubiquitous datamining, by including the device's ability to process their own data locally constitutes an interesting alternative to massively centralized data analysis. Federated learning is a realization of ubiquitous datamining intended to deploy the training of neural net- works on smartphones. We propose here an experimental evaluation of this distributed learning approach on mobile operator data.