Mitigating the performance sacrifice in DP-satisfied federated settings through graph contrastive learning