Federated learning (FL) that enables distributed clients to collaboratively learn a shared statistical model while keeping their training data locally has received great attention recently and can improve privacy and communication efficiency in comparison with traditional centralized machine learning paradigm. However, sensitive information about the training data can still be inferred from model updates shared in FL. Differential privacy (DP) is the state-of-the-art technique to defend against those attacks. The key challenge to achieve DP in FL lies in the adverse impact of DP noise on model accuracy, particularly for deep learning models with large numbers of model parameters. This paper develops a novel differentially-private FL scheme named Fed-SMP that provides client-level DP guarantee while maintaining high model accuracy. To mitigate the impact of privacy protection on model accuracy, Fed-SMP leverages a new technique called Sparsified Model Perturbation (SMP), where local models are sparsified first before being perturbed with additive Gaussian noise. Two sparsification strategies are considered in Fed-SMP: random sparsification and top-$k$ sparsification. We also apply R{\'e}nyi differential privacy to providing a tight analysis for the end-to-end DP guarantee of Fed-SMP and prove the convergence of Fed-SMP with general loss functions. Extensive experiments on real-world datasets are conducted to demonstrate the effectiveness of Fed-SMP in largely improving model accuracy with the same level of DP guarantee and saving communication cost simultaneously.