Research on Heterogeneity Problem of Federated Learning Based on Knowledge Distillation-Translate

With rising computing power, data-driven AI algorithms have become essential in fields like image processing, with significant applications in electricity inspection image detection. However, competition and privacy concerns prevent data sharing between electricity companies, limiting the development of smart inspection technologies. Federated learning offers a solution, enabling companies to collaboratively train a global model without sharing raw data, thus enhancing model performance and data efficiency. When data distribution is inconsistent (non-IID), traditional federated learning faces challenges that affect training efficiency and accuracy. To address this, we propose FedTSD, a federated learning algorithm designed for electricity inspection tasks. By dynamically exchanging teacher-student roles between the global and local models, FedTSD uses a central server to transfer additional knowledge, boosting generalization and reducing overfitting. Experiments on CIFAR-10 and SVHN datasets show that FedTSD outperforms traditional methods, achieving higher accuracy and smoother training, with faster convergence, which is crucial for efficient electricity inspections.