SCOPE: Scalable Training-Free Semantic Coreset Selection for Long-Tailed Federated Learning
SCOPE is a coreset framework for federated learning that filters anomalies and prunes redundant data to handle long-tail data skew. It achieves this by using lightweight scalar metrics shared globally, allowing clients to prune their local datasets adaptively based on a mathematically rigorous consensus from the server. This method ensures efficient communication while improving global model accuracy across diverse, imbalanced datasets.