Nowadays security concerns of computing devices are growing significantly. This is due to ever increasing number of devices connected to the network. In this context, optimising the performance of intrusion detection systems (IDS) is a key research issue to meet demanding requirements on security of complex and large scale networks. Within the IDS systems, attack classification plays an important role. In this work we propose and evaluate the use the generalizing power of neural networks to classify attacks. More precisely, we use multilayer perceptron (MLP) with the back-propagation algorithm and the sigmoidal activation function. The proposed attack classification system is validated and its performance studied through a subset of the DARPA dataset, known as KDD99, which is a public dataset labelled for an IDS and previously processed. We analysed the results corresponding to different configurations, by varying the number of hidden layers and the number of training epochs to obtain a low number of false results. We observed that it is required a large number of training epochs and that by using the entire data set consisting of 31 features the best classification is carried out for the type of Denial-Of-Service and Probe attacks.

Peer Reviewed