Personalized Federated Test-Time Adaptation studies how decentralized models can adapt to changing domains while also remaining robust to severe class imbalance across clients. Existing federated test-time adaptation methods mainly address distribution shift, while class-imbalanced federated learning typically assumes supervised training. We develop pFedBBN, a personalized source-free framework that jointly tackles both problems. At each client, class-wise adaptive normalization uses pseudo-labels to build more balanced feature statistics, while confidence-filtered teacher-student distillation supports controlled unsupervised adaptation. Clients share only batch-normalization statistics, which are used to estimate domain similarity and construct personalized aggregation. Experiments on CIFAR-10-C and CIFAR-100-C show improved robustness and minority-class performance, particularly under highly imbalanced non-IID distributions. A preprint of this work is available on arXiv.
Team: Md Akil Raihan Iftee, Syed Md. Ahnaf Hasan, Mir Sazzat Hossain, Amin Ahsan Ali, A K M Mahbubur Rahman, Rakibul Hasan Rajib, Sajib Mistry, Monowar Bhuyan
