Abstract: Aiming at the problems of low convergence accuracy and easy early convergence of standard GWO, an improved grey wolf optimization algorithm GPGWO was proposed. Firstly, in order to evenly distribute wolves in the search space, the Latin hypercube sampling and reverse learning were combined to initialize the population position. Then, the diversity of wolves in the iterative calculation process was increased, and grey wolves were divided into different types of individuals and different position updating strategies were used. Finally, the alpha Wolf is forced to leave its original location by a random Levy flight. By comparing GPGWO with three improved GWO algorithms on 10 widely used benchmark functions, the simulation results show that GPGWO has certain advantages in optimization. Then, GPGWO is applied in the feature selection scenario of intrusion detection. By combining with different classifiers, GPGWO forms a feature selection algorithm to achieve dimension reduction processing of high-dimensional data sets. Through experiments on intrusion detection data sets, it is proved that the algorithm can retain the optimal feature subset and obtain the best detection effect with only part of the features.