Mejora de la privacidad en Sistemas Inteligentes de Transporte (ITS) mediante aprendizaje distribuido: caso de estudio sistema de control de emisiones en la zona urbana de Cuenca

Abstract

Currently, the world is transitioning towards an interconnected environment, which has increased the availability of large volumes of data and highlighted the importance of protecting privacy. In the realm of mobility, ITS are a significant example of how large amounts of data are generated in interconnected networks. This thesis proposes an architecture based on machine learning techniques, specifically FL, to enhance the privacy of ITS users, with a focus on reducing CO2 emissions in Cuenca, Ecuador. The methodology employed allows data to remain on users’ local devices, sharing only model parameters, thus ensuring greater privacy protection. The use of various machine learning algorithms facilitates data segmentation and classification, thereby reducing vehicle wait times and contributing to decreased CO2 emissions. The results obtained show that FL consumes more resources compared to RL. The average CPU usage in FL is significantly higher, with values close to 41 %, in contrast to 16.69 % in RL. Additionally, FL uses 8.52 % more RAM than RL. This indicates that, while FL offers better privacy protection, it can impact system efficiency. Furthermore, the results reveal that in the context of differential privacy, a higher value of ϵ is associated with lower privacy. Therefore, finding an optimal ϵ value is crucial to balance privacy protection with system efficiency in Federated Learning.