Energy- and AoI-Aware Hierarchical Personalized Federated Learning for Distributed Edge Systems With Selective LLM Module Exchange

초록

Distributed edge systems require federated learning (FL) frameworks that can jointly support timely model updates, user-level personalization, and communication-efficient coordination. However, conventional hierarchical FL (HFL) methods mainly focus on reducing communication overhead and often struggle to maintain model freshness and local adaptability in multitier environments with heterogeneous data and resource constraints. To address these challenges, we propose an energy- and age-of-information (AoI)-aware hierarchical personalized FL (EA-HPFL) framework. The proposed method introduces two key ideas: 1) selective module exchange, in which clients communicate only lightweight adapter, head, or other trainable submodule updates instead of full-model parameters; and 2) deterministic AoI- and energy-aware aggregation, in which freshness and energy indicators guide client- and edge-level aggregation. This design enables efficient hierarchical coordination while preserving local personalization under non-IID data distributions. Experimental results on wearable stress and affect detection (WESAD), human-activity recognition (HAR), and Amazon-EN show that EA-HPFL provides a favorable tradeoff among communication efficiency, update freshness, and personalized performance across diverse edge settings. In particular, the proposed framework maintains competitive accuracy while substantially reducing the overhead associated with full-model exchange, and the revised TinyLlama-based Amazon-EN evaluation further confirms its effectiveness in a large language model (LLM)-oriented setting. These results demonstrate the practical value of AoI-aware coordination and selective submodule exchange for scalable and adaptive FL in real-time intelligent edge systems.

키워드