Reproducible Implementation of an Attention-Enhanced Meta-Learning Network for Personalized Federated Learning with Heterogeneous Client Data

This protocol describes the implementation and evaluation of attention-enhanced meta-learning personalized federated learning (AEML-PFL) for heterogeneous image-classification clients. The custom PyTorch simulation uses 50 clients, synchronous aggregation, and five independent random-seed runs. It evaluates Fashion-MNIST, Federated Extended MNIST (FEMNIST), and the Canadian Institute for Advanced Research datasets CIFAR-10 and CIFAR-100. Client data are not independent and identically distributed (non-IID). FEMNIST uses the same 50-writer cohort of 18,642 images across all five seeds; the other datasets use Dirichlet-based label partitions. Personalized test accuracies are 93.6%, 86.0%, 74.3%, and 47.1%, respectively, after checkpoint selection by validation accuracy. On CIFAR-10, the mean accuracy of the five worst-performing clients is 62.7%, and the across-client standard deviation is 6.2 percentage points. The number of rounds to 95% of each method's own peak validation accuracy is 166 for AEML-PFL and 214 for federated averaging. The protocol provides a workflow for examining personalized performance under label and sample-count heterogeneity.