Resource-Aware Federated Transfer Learning with Adaptive Model Scaling and Channel Attention

The proliferation of heterogeneous edge devices creates opportunities for distributed intelligence but also introduces computational and communication challenges in Federated Learning (FL), with energy consumption emerging as a critical bottleneck for resource-constrained edge devices. Federated Transfer Learning (FTL) mitigates these issues by enabling cross-domain knowledge transfer, improving convergence, and reducing communication overhead. However, existing FTL approaches largely overlook device heterogeneity and resource constraints, leading to suboptimal efficiency and limited applicability in real-world edge environments. To address this gap, we present Resource-Aware Federated Transfer Learning (RA- FTL), a framework that adapts both model architecture and resource utilization to heterogeneous client capabilities. RA-FTL integrates two mechanisms: (1) capability-aware model scaling, which dynamically adjusts trainable parameters according to device profiles, and (2) adaptive channel attention, which sparsifies feature maps based on resource constraints. Experiments across diverse heterogeneous client settings show that the RA- FTL framework achieves 2–6× lower communication overhead, over 7% faster training, and up to 69% reduction in energy consumption compared to FTL and baseline FL methods, while maintaining competitive model performance.