Federated Learning (FL) allows models to be trained across decentralized clients without exchanging raw data, enhancing privacy. Despite this advantage, privacy concerns persist because local gradients shared with an untrusted aggregation server could potentially leak sensitive information. To address this, we introduce a dropout-resilient and privacy-preserving framework for federated learning via lightweight masking (DRPFed), which leverages secure masking and a trusted third party (TTP). The approach utilizes the Diffie-Hellman key exchange protocol to create shared secret keys between clients and the TTP, which are then used to generate masks for obscuring local gradients before they are sent to the server. To guarantee accurate aggregation, the TTP provides the final client with a compensatory mask, ensuring that the combined masks cancel out. Additionally, if a client disconnects, the TTP reallocates the missing mask among the remaining active clients to preserve aggregation correctness. Experimental evaluations demonstrate that, unlike the standard FedAvg, our method maintains model accuracy while effectively handling client dropouts. The proposed solution successfully protects gradient privacy against honest-but-curious servers and malicious clients, all while upholding the reliability of federated model training.