FedCVESA: Taking Away Training Data in Federated Learning via Correlation Value Encoding and Segmented Aggregation