FedBM: Stealing knowledge from pre-trained language models for heterogeneous federated learning.

Area of Science:

• Medical Image Computing
• Artificial Intelligence
• Machine Learning

Background:

• Federated learning (FL) enables collaborative model training without data privacy leakage.
• Data heterogeneity in FL causes local learning bias, degrading overall system performance.
• Existing methods struggle to mitigate bias in heterogeneous federated learning environments.

Purpose of the Study:

• To introduce a novel framework, Federated Bias eliMinating (FedBM), to address local learning bias in heterogeneous FL.
• To improve the performance and robustness of federated learning systems dealing with diverse data distributions.

Main Methods:

• Linguistic Knowledge-based Classifier Construction (LKCC): Utilizes class concepts, prompts, and pre-trained language models (PLMs) to create concept embeddings and estimate latent class distributions.
• Concept-guided Global Distribution Estimation (CGDE): Employs concept embeddings to train a conditional generator for producing pseudo data to calibrate local feature extractors.
• Frozen classifier and generator calibration techniques are employed to prevent bias during local training.

Main Results:

• FedBM framework effectively eliminates local learning bias in heterogeneous federated learning.
• Experimental results on public datasets demonstrate superior performance compared to state-of-the-art methods.
• Ablation studies confirm the significant contribution of both LKCC and CGDE modules.

Conclusions:

• FedBM offers a robust solution for mitigating data heterogeneity challenges in federated learning for medical imaging.
• The proposed approach enhances classifier and feature extractor performance by addressing local learning bias.
• The framework shows promise for advancing privacy-preserving and high-performance federated learning applications.