Most of the neural networks-based intrusion detection systems (IDS) examine all data features to detect intrusion or misuse patterns. Some of the features may be redundant or contribute little (if anything) to the detection process. That is why the purpose of this study is to identify important KDD features which will be used to train a neural network (NN), in order to best classify and detect attacks. Four NNs were studied: Modular, Recurrent, Principal Component Analysis (PCA), and Time-Lag recurrent (TLR) NNs. We investigated the performance of combining the Fisher's filter used as a feature selection technique, with one of the previously cited NNs. Our simulations show that using Fisher's filter improves largely the performance of the four considered NNs in terms of detection rate, attack classification, and computational time.