Statistical prototype exchange and alignment for personalized federated learning

Federated learning within edge–cloud systems frequently encounters double jeopardy from non-independent and identically distributed (non-IID) conditions when combined with label noise from unreliable data acquisition. Existing approaches struggle to address this compound challenge. Conventional aggregation methods fail to facilitate robust learning in the presence of inconsistent labeling. Furthermore, prototype-based frameworks exhibit significant limitations under non-IID conditions. These frameworks tend to become biased toward representations of majority classes, leading to overfitting, while ignoring minority classes. This results in a failure to capture the optimal distribution. In this paper, we propose a framework called Statistical Exchange of Prototypes with Local and Global Alignment for Personalized Federated Learning (SPA-PFL), which leverages statistics-based aggregation that dynamically reweights client contributions based on their reliability. It prioritizes updates with a lower statistical variance in order to maintain global model stability. Moreover, to address representation bias in non-IID scenarios, we design a separation loss that uses global prototypes to regulate the latent space of absent classes, thereby preventing local models from collapsing into biased representation. Experiment results on standard benchmarks and real-world medical datasets demonstrate the effectiveness of SPA-PFL. Specifically, SPA-PFL achieves an average accuracy improvement of 9.21% with the CIFAR-100 dataset, and 5.12% with the International Skin Imaging Collaboration 2019 dataset, compared to existing state-of-the-art baselines. Our framework achieves significant communications efficiency, reducing data transmission by 97.94% on average, relative to standard model-averaging approaches, while ensuring robust convergence despite severe label noise in a non-IID setting.