FedTHQ: Tensor-Assisted Heterogeneous Model With Quality-Based Aggregation for Federated Learning Integrated IoT

The extensive deployment of the Internet of Things (IoT) devices has highlighted significant challenges related to data privacy, security, and communication bandwidth. Federated Learning (FL), as a promising distributed machine learning paradigm, has been integrated with IoT to enhance data privacy, reduce latency and improve learning quality. However, the integration of FL with IoT still faces inevitably challenges due to system heterogeneity and data heterogeneity. We propose a Federated Learning framework based on Tensor-assisted Heterogeneous model with Quality-based aggregation (FedTHQ). FedTHQ can attain a global model with high accuracy and rapid convergence in heterogeneous IoT scenario. FedTHQ mainly includes following two sub-processes. Firstly, the tensor-based heterogeneous model split scheme (TBFL) is proposed to develop a high-performance small model, which is applied in the FL initialization phase. In TBFL, each client reconstructs the local model according to the binary parameter groups. Secondly, we propose quality-based aggregation scheme (QBFL) to assign appropriate weights to each local model considering the heterogeneity. QBFL is applied in the aggregation phase. Experimental results demonstrate the effectiveness and superiority of the proposed schemes. FedTHQ outperforms the benchmarks in global model accuracy, while achieving rapid convergence.