<p>As a distributed learning paradigm, federated learning (FL) allows users to hold datasets in a local environment while training collaboratively. However, there are still problems such as data heterogeneity and model security in federated learning. Adversaries can launch model inference attacks or poisoning attacks to local models. To address the issues above, we present a clustered federated learning framework RPCFL based on secure multiparty computation (SMPC), which aims to achieve dynamic clustering and robust aggregation within groups while protecting privacy. We propose a secure centrality evaluation protocol for dynamically adjusting client clustering to compensate for possible errors in one-shot clustering and adapt to the dynamic changes in data distribution. Finally, we achieve the RPCFL scheme and assess it on two benchmark datasets. The research results indicate that this scheme maintains high performance when dealing with malicious attackers in clustered federated learning. Even when 50% of the clients were malicious, we achieved over 3% higher accuracy than the baseline solution across all the tested datasets.</p>