Geometry-Aware Aggregation for Federated Learning: Down-Weighting Label-Corrupted Clients via Latent Class Prototypes
Abstract
FedAvg primarily weights client updates based on the volume of data for each participant, without considering the reliability of local labels. This approach can be problematic when a large client has unreliable labels, as its contribution remains significant only because of the number of samples it contains. To address this issue, we introduce Geometric Similarity Weighted Aggregation (GSWA), which combines the number of samples with information derived from the geometric properties of the clients’ latent representations. Each client is represented by class-conditional prototypes generated from its latent features. For each class, these prototypes are evaluated against a coordinate-by-coordinate median reference created from the participating clients. The greater the differences between the prototypes, the lower the similarity factors. In this study, we evaluated GSWA on a federated simulation using the scikit-learn digits dataset, distributed across six clients via a Dirichlet partition to create non-independent and non-identically distributed (non-IID) data. When the labels of the most important client were randomly reassigned, GSWA achieved a final accuracy of 96.13% ± 1.35%, compared to 85.82% ± 9.43% for FedAvg over five random initializations. With correct labels, both methods produced similar average accuracies: 97.20% for GSWA and 97.29% for FedAvg. In the evaluated context, GSWA significantly reduced the influence of the corrupted client while maintaining performance close to that of FedAvg when the labels were correct.