Fiche de MESTOUKIRDI Mohamad

Titre : Reliable and Communication-Efficient Federated Learning for Future Intelligent Edge Networks
Date : 4 décembre 2023
Heure: 14:00
Résumé : The focus of the thesis is advancing efficient and robust FL to enable embedded intelligence in 6G networks and innovative edge. A major challenge is statistical heterogeneity from divergent user data distributions, where conventional one-model-fits-all performs poorly. We propose user-centric rules yielding distinct models tailored to each user objectives. Users are clustered by similarity to enable collaborative training of shared models, reducing communication overhead. Another focus is integrating remote IoT devices into intelligent edges using UAVs as FL orchestrators. We jointly optimize UAV trajectories and device scheduling based on a proposed proxy metric of training performance. Finally, we examine the communication overhead of FL wireless networks. To face this, we leverage sparse random networks to approximate target networks with over-parameterized random networks via pruning rather than direct optimization, leading to substantial improvements in communication efficiency.
Lieu : Eurecom