Joint Communication and Over-the-Air Computation for Semi-Federated Learning Towards Scalable AI in Computing-Heterogeneous IoT Systems

TL;DR

This work introduces SemiFL, a hybrid centralized/ federated learning framework designed for computing-heterogeneous IoT environments to enable universal participation. It proposes a next-generation multiple access (NGMA) scheme that unifies NOMA-based raw-data uploads and AirComp-based gradient aggregation, enabling simultaneous data and gradient transmissions with SIC and receive beamforming. The framework is augmented with STAR-RIS and SWIPT modalities to tackle channel diversity and energy constraints, demonstrating superior learning performance over baselines on MNIST with faster convergence and higher accuracy. The results highlight the potential for scalable, efficient edge learning in diverse IoT networks, while open issues point to theory, prototyping, and regulatory challenges for OTA model training. Overall, SemiFL presents a practical path toward integrated computation and communication for large-scale, heterogeneous IoT deployments.

Abstract

The proliferation of Internet of Things (IoT) systems demands scalable artificial intelligence (AI) solutions that can operate in computing-heterogeneous environments with diverse hardware capabilities and non-independent and identically distributed data. This paper proposes a semi-federated learning (SemiFL) framework that integrates centralized learning (CL) and federated learning (FL) to enable efficient model training across heterogeneous IoT devices. In SemiFL, only devices with sufficient computational resources are designated for local model training (referred to as FL users), while the remaining devices transmit raw data to a base station (BS) for remote computation (referred to as CL users). This collaborative computing framework enables all IoT devices to participate in global model training, regardless of their computational capabilities and data distributions. Furthermore, to alleviate radio resource scarcity of SemiFL, we propose a joint communication and over-the-air computation design that unifies wireless transmission and model aggregation. This approach reduces latency and enhances spectrum efficiency by allowing simultaneous communication and computation over the air. Furthermore, we design a transceiver architecture that integrates receive beamforming and successive interference cancellation to mitigate multi-user interference, ensuring reliable aggregation at the BS. Subsequently, two scenarios are examined to assess the efficacy of SemiFL within IoT systems. Simulation results demonstrate that the proposed next-generation multiple access (NGMA)-based SemiFL outperforms the fixed beamforming-based FL by 8% and the AirComp-based FL by 6.4%, demonstrating its superior learning efficiency and convergence performance.

Figures (6)

• Figure 1: The first row illustrates three typical ML paradigms: CL, on-device learning, and FL. The second row demonstrates three common uplink transmission schemes for wireless FL, including OMA, NOMA, and AirComp. Specifically, CL is often deployed in the cloud-based system that has rich datasets and powerful computing capabilities. When implemented in IoT networks, CL entails key problems, such as high communication overhead, privacy risk, and inference delay. On-device learning allows all raw data to be processed on local devices. However, implementing on-device learning among low-end IoT devices faces issues such as insufficient computing capability, limited memory capacity, constrained energy supply. Although FL has presented huge potential in various intelligent IoT applications, it also faces challenges, such as unbalanced data distribution and computing-heterogeneous IoT devices.
• Figure 2: The proposed SemiFL framework and its learning process.
• Figure 3: Schematic diagram of the proposed NGMA-based SemiFL framework and its transceiver design. The top half illustrates the key steps of an NGMA-based SemiFL framework in one communication round (i.e., Steps 3-6). The bottom half provides the corresponding transceiver design and the procedure of signal processing.
• Figure 4: A case of SemiFL in STAR-RIS-aided IoT systems. The STAR-RIS enhances data accessibility and manages interference, allowing for more reliable gradient updates and dataset uploads, which contribute to the overall learning performance of SemiFL.
• Figure 5: An illustration of the SWIPT-enabled SemiFL framework and its transceiver design. These IoT devices harvest energy from the BS to power their operations, including data collection, local model training, and uplink transmission.