Differential Privacy: Exploring Federated Learning Privacy Issue to Improve Mobility Quality

The increasing number of vehicles within cities has been stimulating intense research for smart tools to avoid traffic jams and improve mobility quality. To get the most accurate data, urban mobility applications consume data generated by users as they explain the mobility pattern of a studied region, entailing the meantime vulnerability of privacy of these latter. Thus, new techniques are considered to share the minimum information possible to train models and keep privacy. Federated Learning (FL) is one approach that exploits this concept, sharing only the gradients produced by local models trained at each device. Nonetheless, information can be learned during the training, exposing vulnerable content. In this context, this study aimed to validate the FL in a given application and check the data leakage in this decentralized training architecture through the differential privacy (DP) method, removing potential clients that make the architecture vulnerable. Finally, we ended up having a final FL model with 71% of accuracy and removing two processing units from the training process.