This paper presents a privacy-enhanced federated learning (FL) framework that safeguards client data privacy during collaborative model training by integrating local differential privacy (LDP) and secure multi-party computation (MPC). Traditional FL methods are vulnerable to privacy inference attacks, where malicious entities can extract sensitive information from shared model parameters and gradients. To mitigate these risks, the proposed framework first applies LDP to add Laplace noise to client gradients and then leverages MPC to securely aggregate noisy gradients without revealing individual gradients. This dual-layer approach effectively enhances privacy protection while maintaining high model performance and efficiency, as demonstrated by extensive experiments that show minimal accuracy and efficiency loss. Our results indicate that the combined use of LDP and MPC offers superior privacy guarantees compared to using either method independently, especially in scenarios susceptible to privacy inference attacks. By ensuring robust privacy throughout the FL training cycle, this framework provides a reliable solution for privacy-critical federated learning applications.