Federated Data Engineering and Learning for Edge Intelligence Systems

Afsaneh Mahanipour at the University of Kentucky has completed a doctoral dissertation outlining a new approach to federated learning for edge intelligence systems. The research addresses significant barriers preventing the deployment of deep learning in distributed environments, including limited computation, communication bandwidth, and energy resources. The work identifies data preprocessing as a critical bottleneck and introduces a comprehensive framework for federated data engineering. This includes novel methods for federated multi-label feature selection that utilize information theory and reinforcement learning to reduce dimensionality without sacrificing predictive performance. Additionally, the dissertation proposes Federated Reprogramming Knowledge Distillation (FedRD) to allow lightweight client-side models to benefit from powerful foundation models kept centrally. Beyond preprocessing, the study presents an embedded dynamic sparse federated feature selection method that integrates feature selection with model training. Extensive evaluations on real-world datasets across healthcare, IoT, and multimodal learning tasks demonstrate that these proposed methods significantly reduce communication overhead and inference latency. The findings contribute a unified and scalable framework intended to advance the deployment of privacy-preserving AI in real-world cyber-physical environments.