<p>To enhance model performance while ensuring data privacy, federated learning has become a key approach in distributed training, allowing users to collaboratively train a more accurate global model by sharing model gradients rather than raw data. However, since attackers may infer sensitive information from the local weights, encrypting them is vital for privacy protection. Moreover, issues such as Byzantine attacks, where malicious users upload deceptive the local weights, can damage training outcomes and accuracy. Tackling these challenges is critical for the integrity of federated learning systems. Current methods seldom simultaneously address data privacy, robustness, and efficiency. We propose a novel Privacy-preserving and Robust Federated Learning Scheme (PRFLS) with High Efficiency through Ciphertext Filtering to address these gaps. We first introduce a lightweight encryption protocol that safeguards users’ local weights, enabling secure aggregation with high computational efficiency. Then, we develop a weight filtering algorithm under ciphertext to eliminate abnormal local weights while preserving users weight privacy. This algorithm significantly improves the accuracy and robustness of the global model, particularly when local weights deviate considerably. Lastly, we present the security proof to prove that the proposed scheme can meet the privacy requirements of users’local weights. We also evaluate its performance over MNIST and Chest X-ray datasets and compare it with existing related works. The results show that this scheme is an efficient solution to protect the privacy of local weights while maintaining high model accuracy.</p>