The growing volume of data on the computer network led to increasing the challenges for intrusion detection systems to deal with high dimensions that contain irrelevant and redundant features. This consumes time and difficulty in detecting the attack correctly, with increasing false alarms rate. This problem can be solved by applying dimensionality reduction. In this paper, a wrapper feature selection model based on Differential Evolution technique is proposed for intrusion detection systems. It reduces the number of features by finding the minimum number of features without effecting on the performance of the system. The main idea is to select some features from 41 features of NSL-KDD datasets using Differential Evolution and evaluate these features by computing the accuracy using Extreme Learning Machine. Differential Evolution is continued until obtaining the minimum number of features that satisfy a high accuracy. The results have shown a better detection rate with reduced false alarm rate in five and binary classification. The proposed system achieved an accuracy of 80.15 % and 87.53 for five and binary classification respectively with a reduction in training and testing time.