FedTCTF: Tensor Completion-Based Federated Learning for Device Heterogeneity

Abstract

With the advancement of Artificial Intelligence of Things (AIoT) systems, edge devices are increasingly capable of participating in Federated Learning (FL) by collaboratively training a global model while preserving data privacy. However, the low bit-width and limited storage capacity of edge devices often degrade global model accuracy, leading traditional FL frameworks to resource waste. To address this issue and enhance edge device utilization, we propose FedTCTF, an innovative framework that integrates optimized tensor completion with a novel traversal-based Layer-Wise Model Partitioning strategy. Our approach tackles both bit-width heterogeneity and capacity constraints through two key innovations: (1) A bit-width-aware tensor completion method using adaptive Tucker and CP decomposition to align parameter precision across devices; (2) A layer-wise dynamic sub-model extraction mechanism that systematically cycles through all parts of the global model via a traversal window. Specifically, the server dynamically extracts sub-models from the global model based on device capacities, with the traversal window advancing layer-by-layer in each training round to extract layers not covered in the previous rounds, ensuring resource-constrained devices evenly train the full model. Experimental results demonstrate FedTCTF’s superiority: It achieves up to 20.77% higher accuracy on CIFAR-10 with ResNet18 compared to state-of-the-art methods, and consistently outperforms all baselines on NLP tasks, including AGNews and StackOverflow with Transformer, particularly excelling when 80% of clients are low-capacity devices. By enabling comprehensive device participation through precision recovery and uniform model coverage, FedTCTF establishes a sustainable learning paradigm for AIoT ecosystems.