DAP-MFL: Distributed AP-Assisted Multilayer Federated Learning for Resource-Constrained IoT Networks With RAW-Slot Management

This paper introduces DAP-MFL, a novel Distributed Access Point-Assisted Multilayer Federated Learning framework tailored for resource-constrained IoT systems. Recognizing the limitations of traditional 4G/5G-based FL deployments in terms of energy consumption and scalability, DAP-MFL strategically leverages an energy-efficient IEEE 802.11ah network protocol to enable more sustainable federated learning implementations. The framework introduces a client–edge–fog–cloud architecture that systematically distributes the aggregation process across multiple network layers, thereby optimizing both computational and communication resources while maintaining the integrity of the learning process. At the core of DAP-MFL are three specialized slot management methods developed to address the unique channel access constraints of IEEE 802.11ah networks: (1) Strict Slot Assignment with Dropping (SSAD), which introduces a strict dropping policy for latency-sensitive scenarios, maximizing efficiency by removing stragglers instantly; (2) Selective Replacement with Gradual Inclusion (SRGI), which prioritizes stability via a unique phased-replacement strategy, gradually integrating nonparticipant stations without disrupting training; and (3) Round Robin (RR), which provides fairness in resource allocation. Extensive evaluations using real-world IoT datasets (AQC and WISDMM) and the benchmark MNIST dataset demonstrate that SSAD achieves the highest latency reduction of up to 97% and energy savings of up to 95% compared to the baseline methods, while SRGI simultaneously offers significant latency reduction of up to 70% and energy efficiency improvements of up to 78% along with enhanced stability and long-term learning without compromising the accuracy compared to the baseline methods. These results validate the efficacy of our framework, showcasing SSAD’s superior efficiency of SSAD and SRGI’s balanced performance of SRGI, making it well-suited for large-scale IoT deployments in smart cities and beyond.