Privacy-preserving Federated Primal-dual Learning for Non-convex and Non-smooth Problems with Model Sparsification