TSFL-SS: Task-Specific Federated Learning for Resource-Limited Edge Devices

The widespread deployment of IoT devices in resource-constrained edge environmentsłcharacterized by limited computational power, storage, and network bandwidthłposes significant challenges for federated learning (FL). Traditional FL frameworks inadequately support frequent model updates and complex computations, while existing methods fail to balance computational efficiency and model performance on resource-limited devices, resulting in inefficient training and prohibitive communication costs. To address these limitations, we propose TSFL-SS, a task-specific federated learning framework that integrates three synergistic modules: (1) an Importance-Aware Sub-model Selection (IASS) module dynamically extracts lightweight task-relevant parameters via one-shot low-rank decomposition, enabling edge devices to bypass full-model computations; (2) a Task-Specific Model Partitioning (TSMP) module decouples models into shared encoders (fixed task-agnostic features) and personalized predictors, minimizing redundant transmissions; and (3) a Semantic-Aware Aggregation (SAP) module aligns client updates by clustering parameters with high semantic similarity, mitigating conflicts from non-IID data distributions. Experiments on ResNet18 demonstrate that TSFL-SS achieves a 26.84% average accuracy improvement over FedAvg across both IID and non-IID settings, while reducing communication overhead by 55.44%. These advancements establish TSFL-SS as an efficient, scalable solution for FL in resource-constrained edge environments.