@article{M31F51DDB,
title = "Personalized Federated Post-Training for Real-time Adaptation to Label Distribution Shifts",
journal = "The Journal of Korean Institute of Communications and Information Sciences",
year = "2025",
issn = "1226-4717",
doi = "10.7840/kics.2025.50.12.1811",
author = "Kyungjin Im, Heewon Park, Miru Kim, Mugon Joe, Minhae Kwon",
keywords = "Federated Learning, Post-training, Personalization, Label Shift",
abstract = "Artificial intelligence systems deployed to individual devices are exposed to shifting label distributions over time, degrading model performance. While post-training has been studied to address this, limited local data and overfitting can cause loss of critical knowledge in the pre-trained model. We propose Fed-AFIR(Federated Adaptation with FIM Regularization), leveraging the Fisher Information Matrix (FIM) to preserve critical parameters while enabling collaboration across devices, consistently outperforming existing approaches under dynamic label distribution and heterogeneous device data conditions."
}