Aiming at the problems of poor training performance and high false positive detection rate of existing intrusion data feature extraction algorithms in small sample environment, a feature extraction algorithm based on fusion residual network is designed. Firstly, the small sample data is preprocessed to reduce the dimension of the input data. Secondly, the residual module stack is used to form a convolutional layer to improve the data training ability of the model and the feature extraction ability of small sample data. Based on the gating mechanism of the short-time memory network model, the flow loss of input data features is adjusted and controlled in real time. Finally, the minimum objective function method is used to optimize the final feature classification results. The experimental results show that the data set detection time of fusion residual network is short, iteration can be completed within 60 times, and the detection accuracy rate for different types of attacks is more than 97%.