Networks intrusion detection systems allow to detect the various attacks. The problem of precision and flexibility makes IDS inefficient. One of the major reasons of such a situation is the misconception of a correct profile. In this paper, a modular architecture is proposed. This architecture is applied to neural models without attribute selection (MAMuM), with selection (MAMuMS) and in deep learning (MAMuMD). Each module is dedicated to detecting a particular type of attack. To improve those systems performances, it is necessary to select the most relevant attribute that will lead to characterize a normal profile or an attack. We have proposed in this paper an intrusion detection system architecture based classification and coupled with to flexibly select attributes using neural networks. We present a new model of attack type choice. We explain some type of attack. We have done a comparative study with others works. The benchmark dataset NSL-KDD and UNSW-NB15 has been used to train, test and evaluate our work. Our architecture coupled with selection and deep learning improves the accuracy of the intrusion detection. We obtained an average detection rate of 97.3% for the NSL dataset and 92.9% for the UNSW dataset.