FedFC: A Personalized Federated Learning based feature representation and classifier combination
Liming Chang · 2024
Federated learning aims to address the challenges posed by data fragmentation and data isolation while ensuring privacy protection in distributed machine learning. In the federated learning setting, each participant collaboratively trains a local model using their local data. These locally trained models are subsequently aggregated at the server node. However, in real-world application environments, the data distribution between clients often varies significantly, leading to suboptimal precision of federated learning models. To mitigate the impact of non-independent uniformly distributed data on model accuracy, we introduce a novel approach. Our method combines common representation learning, achieved through prototype contrastive learning, with personalized classifier collaboration based on similar data distribution. The goal is to learn effective personalized local models. Extensive evaluation results on benchmark datasets, considering various heterogeneous data scenarios, consistently demonstrate the effectiveness of our proposed method.