In the field of Internet of things (IoT) intrusion detection, federated learning has become an effective solution for implementing model weight integration updates. This distributed learning method allows devices to train models locally and transmit updated parameters to a central server for aggregation. However, existing intrusion detection methods based on federated learning still have limitations. In scenarios with non-independent and identically distributed data and heterogeneous client models, the intrusion detection performance of the global model will be severely affected. The significant communication overhead caused by simultaneously transmitting model parameters also hinders the actual deployment of federated learning schemes. To address the aforementioned issues, an efficient IoT intrusion detection method based on semi supervised federated learning is proposed. By utilizing unlabeled public data to enhance the model's understanding of the data, the performance of the client classifier is continuously improved. At the same time, a discriminator module is added to improve the quality of the client's predicted labels, and the combination of hard label strategy and voting mechanism effectively reduces communication overhead. The experimental results show that an accuracy of 86.97% is achieved in non-independent and identically distributed data and heterogeneous client model scenarios, which is superior to typical federated learning methods and achieves lower communication overhead.