Enhancing Test-Time Personalized Federated Learning via Dual Disentangled Factors

Traditional test-time adaptation (TTA) methods are often designed for centralized training models, and many hyperparameters set for these methods do not translate well to personalized federated learning with test-time adaptation (PFL-TTA). We propose FedDualPLPD, a novel PFL-TTA framework that leverages domain-independent latent disentangled factors during model training to guide TTA in the PFL-TTA context. These latent factors-representing underlying attributes such as object shape and background texture-play a key role in influencing predictions. By computing Dual Pseudo-Label Probability Difference (DualPLPD) scores during federated training, our method identifies and prioritizes samples driven by these generalizable factors, enabling robust adaptation during testing. The framework follows a two-stage adaptation strategy: (1) computing the latent disentangled factors during the source-domain federated training phase, and (2) using these domain-independent PLPD scores during the testing phase for effective test-time adaptation. This approach ensures the appropriate use of TTA techniques in the PFL-TTA setting, leading to improved performance and more efficient deployment in distributed, heterogeneous environments.