Abstract: Information is an important asset of an organization. Large amount of information need to be stored and processed in network based computers. The confidentiality, integrity and availability of the system resources have raised the vulnerability of these systems to security threats, attacks and intrusions. One idea is to use a neural network algorithm for detecting intrusions. The neural network algorithms are popular for their ability to ’learn’ the patterns in a given environment and thus can be trained to detect intrusions by recognizing patterns of an intrusion. In this work we perform a comparative study of Multilayer Feed Forward, Elman Back Propagation, Cascaded Forward Back Propagation and Self Organizing Feature Map neural networks based intrusion detection systems. In this study we work on the well structured KDD CUP 99 dataset.