Abstract: Federated learning (FL) provides a learning framework without participants sharing local raw data, but individual privacy is still at risk of disclosure through attacking the trained models.