28 January 2025
08:30 Doctoral defense Room 85 of IC2
Theme
Mobility-Aware Federated Learning in Vehicle Edge Computing for Autonomous Vehicle Applications
Student
Wellington Viana Lobato Junior
Advisor / Teacher
Leandro Aparecido Villas - Co-advisor: Denis Lima do Rosário
Brief summary
"The rapid proliferation of Connected and Autonomous Vehicles (CAVs) highlights the critical need for efficient privacy-preserving data processing frameworks in vehicular networks. This thesis proposes a set of solutions to integrate federated learning (FL) with mobility awareness within Vehicular Edge Computing (VEC) in order to address the unique challenges posed by CAV environments, such as high mobility, dynamic communication conditions, and resource heterogeneity. FL offers a decentralized model training paradigm that preserves privacy by keeping data localized, a crucial feature given the sensitivity of vehicular data. However, its application in VEC is limited by factors such as intermittent connectivity, constrained local computational resources, and data heterogeneity. This thesis presents three main contributions to overcome these challenges: FLEXE, a simulation environment designed to evaluate FL in VEC scenarios by integrating vehicular network simulators with machine learning frameworks, which enables a more accurate and accurate analysis of the data realistic FL application model; GYRFALCON, a semi-asynchronous FL mechanism that dynamically adjusts synchronization points based on real-time connectivity metrics, optimizing aggregation and reducing communication overhead, without compromising model accuracy; MELRO, a multi-model FL assignment algorithm that dynamically allocates models to clients based on link duration, data entropy, and training latency, ensuring efficient use of edge resources and minimizing incomplete tasks. The solutions proposed in this thesis were compared with literature approaches on several performance metrics, considering realistic vehicular mobility scenarios. The presented contributions increase the scalability, robustness, and adaptability of FL for VEC, offering solutions for the development of privacy-preserving systems supporting autonomous vehicle applications. The results demonstrate that the developed solutions are robust to face the challenges of high mobility and resource heterogeneity, promoting the efficiency of FL in VEC environments."
Examination Board
Headlines:
Leandro Aparecido Villas | IC / UNICAMP |
Augusto José Venâncio Neto | UFRN |
Rodolfo Ipolito Meneguette | ICMC / USP |
Juliana Freitag Borin | IC / UNICAMP |
Lucas Francisco Wanner | IC / UNICAMP |
Substitutes:
Luiz Fernando Bittencourt | IC / UNICAMP |
Judy Carolina Guevara Amaya | IC / UNICAMP |
Bruno Yuji Lino Kimura | ICT / UNIFESP |
Maycon Leone Maciel Peixoto | IC/UFBA |