Abstract: Cyber attacks in the Internet of Medical Things (MIOT) can paralyze the healthcare system, interrupt treatment, and pose a great threat to patients' lives. Therefore, cyber attack detection in MoIT is imperative. Medical terminals are rich in scenarios and complex in equipment. Cyber attack targets are scattered to form islands, and the means are diverse and unbalanced. Attack detection with a single model is difficult to work. Therefore, a medical Internet of Things attack detection technology based on a personalized federated model is proposed. A local model is built on the local client based on the meta-learning framework combined with the ResNet algorithm. The training process includes two stages: In the first stage, the client and the central server conduct model "seeking commonality" training. The local model on the client absorbs all parameters of the global model to ensure the generalization ability of the local model. In the second stage, the client and the central server conduct model "keeping differences" training. The local model on the client absorbs some parameters of the global model and uses the local data classification rules to train a personalized proprietary model. Experimental results show that the first stage of seeking commonality training only requires a few iterations to achieve a high detection accuracy, and the second stage of personalized training freezes a few network structures to enable the local model to achieve a high detection accuracy. In addition, by comparing with other federated learning methods, the personalized federated algorithm has good detection performance for local network attacks.