COMMUNICATION-EFFICIENT AND SUSTAINABLE FEDERATED LEARNING FOR EDGE AI APPLICATIONS

Rapid growth in IoT deployments and the need for intelligent systems with real-time and privacy-preserving capabilities have led to the development of Edge Artificial Intelligence (AI). Federated Learning (FL) has been a vital component that facilitated the transition from Cloud-based AI to Edge AI by permitting multiple edge devices to jointly learn a common model without transferring data to a centralized server. The deployment of FL on limited-edge devices faces several challenges such as the communication overhead incurred by exchanging frequent model updates and the environmental cost associated with the training of large models. This chapter addresses the interplay between communication-efficient Federated Learning, model compression, and green computing as complementary solutions for Edge AI systems. The paper starts from basics of Edge AI and the FL process, followed by a critical review of communication-efficient approaches such as gradient sparsification, quantization, adaptive client sampling, and asynchronous aggregation. This is followed by model compression methods such as pruning, quantization, knowledge distillation, and efficient architecture designs, which both minimize computation and communication. Sustainable and green computing concepts are then considered along with their adoption in FL pipelines. Practical use cases in healthcare, smart cities, Industrial Internet of Things, and precision farming are then presented. Finally, important challenges like device heterogeneity, security, and scalability are identified and discussed, followed by future directions of research such as edge LLMs, green FL, and 6G-based intelligence.