Drone communication is currently a hot topic of research, and the use of drones can easily set up communication networks in areas with complex terrain or areas subject to disasters and has broad application prospects. One of the many challenges currently facing drone communication is the communication security issue. Drone communication networks generally use software defined network (SDN) architectures, and SDN controllers can provide reliable data forwarding control for drone communication networks, but they are also highly susceptible to attacks and pose serious security threats to drone networks. In order to solve the security problem, this paper proposes an intrusion detection model that can reach the convergence state quickly. The model consists of a deep auto-encoder (DAE), a convolutional neural network (CNN), and an attention mechanism. DAE is used to reduce the original data dimensionality and improve the training efficiency, CNN is used to extract the data features, the attention mechanism is used to enhance the important features of the data, and finally the traffic is detected and classified. We conduct tests using the InSDN dataset, which is collected from an SDN environment and is able to verify the effectiveness of the model on SDN traffic. The experiments utilize the Tensorflow framework to build a deep learning model structure, which is run on the Jupyter Notebook platform in the Anaconda environment. Compared with the CNN model, the LSTM model, and the CNN+LSTM hybrid model, the accuracy of this model in binary classification experiments is 99.7%, which is about 0.6% higher than other comparison models. The accuracy of the model in the multiclassification experiment is 95.5%, which is about 3% higher than other comparison models. Additionally, it only needs 20 to 30 iterations to converge, which is only one-third of other models. The experiment proves that the model has fast convergence speed and high precision and is an effective detection method.