To address the limitations of incomplete feature extraction and the issues of data silos and privacy leakage in traditional centralized intrusion detection systems, an intrusion detection method based on federated learning and spatio-temporal feature fusion was proposed. Convolutional neural networks and long short-term memory networks were used to extract temporal and spatial features respectively. These extracted features were then concatenated in parallel to generate fused features. A multi-head attention mechanism was employed to identify critical characteristics within the network traffic data, followed by training through bidirectional gated recurrent units and final classification via Softmax function. During the model training process, in order to prevent privacy leakage, the inherent characteristics of federated learning were leveraged to enable data to remain local for neural network model training. Experimental results demonstrated that the proposed model achieved accuracy rates of 99.00%, 97.64%, and 75.28% on the CIC-IDS2018, NSL-KDD, and UNSW-NB15 datasets, respectively.