DictPFL: Efficient and Private Federated Learning on Encrypted Gradients