A Federated Learning Framework for Optimizing Edge Computing with Semantic Offloading

Traditional real-time edge tasks, such as monitoring and control, often rely on centralized processing, which introduces single points of failure, high communication costs, and delays, particularly in latency-critical environments. Federated Learning (FL) offers a distributed alternative by enabling collaborative model development without the need to upload local data, thereby reducing the communication overhead and enhancing privacy. However, the limited computational resources of edge devices constrain their ability to store extensive data and run complex deep learning models. To address these challenges, we propose a novel hybrid framework that integrates FL to optimize real-time task execution and decision making. Our architecture consists of three layers: edge devices, regional edge servers (RES), and cloud servers (CS). The RESs, in collaboration with CS, train high-capacity models for themselves and lightweight models for edge deployment, enabling efficient initial data analysis on the edge. To address the sub-optimal performance of lightweight models in complex scenarios, we introduce a semantic offloading mechanism based on uncertainty estimation. When the uncertainty of the model exceeds a predefined threshold, the data are dynamically offloaded to the RES, and high capacity models enhance decision making for these complex cases. This collaborative strategy ensures that most data are processed locally with minimal latency, whereas complex or ambiguous instances benefit from the RES’s superior computational capabilities. Experimental results demonstrate the effectiveness of the proposed method, achieving task execution accuracy improvements of 2.59% to 5.68% over using only lightweight models, closely matching centralized computing systems. Additionally, the proposed method realizes bandwidth savings of 58.73% to 86.86% compared to relying solely on sophisticated RES models across diverse scenarios.