In large data sets data pre-processing always has been the most essential data processing stages. Sampling and using small volumes of data has been an integrated part of data pre-processing to decrease training errors and increase speed of learning. In this study, instead of sampling from all data and using small parts of them, a method has been proposed to not only benefit from sampling but all data be used during training process. In this way, outliers would be detected and even used in completely different way. Using artificial neural networks, new features for instances will be built and the problem of intrusion detection will be mapped as a 10- feature problem. In fact, such a classification is for feature creation and as features in new problem only have discrete values, in final classification decision tree will be used. The results of proposed method on KDDCUP’99 datasets and Cambridge datasets show that this has improved classification in many classes dramatically.