The challenge for artificial intelligence (AI)-based biomedical systems is to ensure privacy and security when sharing sensitive medical data. While collaborative learning offers a solution by allowing model exchange and enabling distributed data analysis, the risk of model poisoning, where malicious participants inject false or manipulated model updates, can lead to potential degradation of overall learning performance. To solve this issue, we proposed a Client Selection Secure Collaborative Learning (CSSCL) algorithm to ensure robust collaborative learning in a hybrid real-time AI-enabled biomedical system while preserving data and model privacy. Specifically, we used Minkowsky similarity as a metric to assess the quality of local models, filter out the manipulated models, and exclude them from the aggregation process. The qualified models related to the corresponding clients were selected, resulting in a reliable aggregated model. In the empirical evaluation of the proposed CSSCL algorithm, the performance metrics showed a remarkable improvement in the final training accuracy. In particular, the proposed method significantly increased the final training accuracy, ranging from 2 to 30% for attack scenarios from 20 to 50%. In addition, the proposed method was evaluated in a specific application scenario, namely pneumonia detection, and provided comparably favorable results. The experiment results showed that the CSSCL-based method had notable potential in various application scenarios. The result of this study presented a significant contribution to medical diagnostic systems for future research.