FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning