Heterogeneous Wireless Federated Learning Framework via Over-the-Air Computation

The vision of sixth-generation networks, powered by artificial intelligence (AI) and edge computing, is to establish a ubiquitous and immersive wireless communication infrastructure that facilitates the transition to the advanced Artificial Intelligence of Things (AIoT) paradigm. To realize this vision, wireless federated learning (WFL) has emerged as a promising approach for distributing AI computations across diverse endpoints while ensuring data privacy within the wireless network. As the exploration of access entities expands, AIoT systems encounter challenges due to the multi-heterogeneity of devices, including diverse computation capabilities and data distribution. To address these challenges, we propose a novel synchronous/asynchronous WFL mechanism with over-the-air computation, which simultaneously ensures time efficiency, communication efficiency, and performance accuracy. Specifically, this framework leverages both global and local momentum to accelerate convergence. Then, a masking matrix is introduced to mitigate heavy noise, balance communication overhead with training loss, and reduce the mean squared error in model aggregation after optimal beamforming design. To further improve performance in such heterogeneous settings, a fair-weighted aggregation method is employed to tackle the biased device selection problem in asynchronous aggregation, especially in the non-independent and identically distributed (i.i.d.) case. Finally, the proposed framework is validated by both i.i.d. and non-i.i.d. MNIST datasets, with extensive numerical results demonstrating fast convergence and improved aggregation performance under varying computational diversity.