Fed-NL: A Federated Learning Approach to Suppress Noise in Participant Datasets to Reduce Communication Rounds for Convergence