Latency-efficient edge intelligence in IoT networks is crucial to support real-time decision-making, low-power data processing, and autonomous operations across smart environments. Knowledge distillation offers a promising direction to compress complex AI models into lighter versions suitable for resource-constrained edge devices. However, existing cloud-based and centralized learning approaches introduce high communication overhead, increased latency, and privacy risks due to frequent data transfers and dependence on remote servers. Traditional distributed learning solutions also struggle with heterogeneous device capacities and model degradation, leading to reduced inference accuracy and slow response time. This paper proposes a Federated Edge Knowledge Distillation (FEKD) framework that leverages distributed teacher–student learning to minimize computational load at the device level, while maintaining global model efficiency. The cloud-based teacher model distills soft knowledge to lightweight student models deployed on edge nodes, enabling reduced model complexity, faster inference, and secure local adaptation without raw data sharing. The proposed method supports latency-sensitive IoT applications such as real-time traffic prediction, healthcare monitoring, and industrial automation, ensuring robust decision-making even under bandwidth limitations. Experimental findings confirm that FEKD decreases latency by up to 40%, improves model accuracy by 12%, and significantly reduces communication overhead while preserving energy efficiency across heterogeneous IoT networks. The proposed method reduced latency between (60–75 ms), model accuracy improvement (92–97%), communication overhead (6–7 MB), energy consumption by 18 mJ, and computational distribution of 50%.
Cui et al. (2026) studied this question.