Toward Cleansing Backdoored Neural Networks in Federated Learning