This paper considers the feature selection problem for data classification in the absence of data labels. It first proposes an unsupervised feature selection algorithm, which is an enhancement over the Laplacian score method, named an Extended Laplacian score, EL in short. Specifically, two main phases are involved in EL to complete the selection procedures. In the first phase, the Laplacian score algorithm is applied to select the features that have the best locality preserving power. In the second phase, EL proposes a Redundancy Penalization (RP) technique based on mutual information to eliminate the redundancy among the selected features. This technique is an enhancement over Battiti's MIFS. It does not require a user-defined parameter such as beta to complete the selection processes of the candidate feature set as it is required in MIFS. After tackling the feature selection problem, the final selected subset is then used to build an Intrusion Detection System. The effectiveness and the feasibility of the proposed detection system are evaluated using three well-known intrusion detection datasets: KDD Cup 99, NSL-KDD and Kyoto 2006+ dataset. The evaluation results confirm that our feature selection approach performs better than the Laplacian score method in terms of classification accuracy.