<p>The goal of intrusion detection is to prevent potential security threats by analysing activities in a network or system and identifying abnormal or malicious behaviours in a timely manner. However, the amount of data in many systems or devices is relatively limited, and the data distribution of different devices may differ significantly, which makes the training data from a single device insufficient to support effective model training. If the data from all devices is uploaded to a central server for centralised learning, it will not only face the pressure of computation and storage, but may also trigger the risk of privacy leakage. In addition, the uneven amount of data may lead to the weak contribution of certain devices to the model. Therefore, this paper proposes FedAware, a federated learning-based intrusion detection method for IoT, which combines a fractal temporal shrinkage encoder with an OC-SVDD single-class classifier to improve the anomaly detection capability. In addition, this paper proposes the ImbalMSE algorithm, which improves the global performance by considering the data imbalance characteristic of IoT device scenarios and the MSE performance difference of local models, and comprehensively ensures that the client with larger amount of data and the better-performing model contribute more to the global model. Experimental results show that FedAware significantly improves the accuracy and robustness of intrusion detection on the N-BaIoT dataset.</p>