FedHLC: A Novel Federated Learning Algorithm Targeting Heterogeneous and Long-Tailed Data for Efficient Image Classification in Consumer Electronics

Zhiguo Qu, Zhiwei Liang

Abstract: Federated learning (FL) is an effective technique for image classification in consumer electronics. This paper proposes a new FL algorithm called FedHLC to address heterogeneous and long-tailed data. Its architecture comprises a feature extractor and a classifier. The training process of FedHLC is divided into two distinct stages. In the first stage, it focuses on training feature extractors on the client side and conducts feature representation learning. This approach develops a robust and generalizable representation for digital image data. The second stage involves retraining the classifier on the server side with generated virtual features. This step not only safeguards client privacy but also effectively mitigates model bias towards tail categories. In addition, FedHLC incorporates a novel balancing factor that dynamically adjusts the influence of regularization term. It allows a flexible focus shift between global objectives and local objectives. The simulation experiments on benchmark datasets demonstrate that FedHLC outperforms the baseline algorithms including CReFF, FedAvg, FedProx and FedNova in terms of accuracy when dealing with heterogeneous and long-tailed data. Furthermore, FedHLC can not only achieve good convergence but also attain an accuracy peak of 89.24%, marking a substantial advancement in the field of FL for image classification in consumer electronics. The code is available at https://github.com/Kiritoliang/FedHLC.

External IDs:dblp:journals/tce/QuL24