Training Function Stability in Anomaly Intrusion Detection based Deep Learning
Khadidja Bennaceur, Zakaria Sahraoui, Abdenour Labed, Mohamed Ahmed‐Nacer · 2019
Deep Learning is a promising model used in both supervised and unsupervised classification as well as in feature extraction. It is applied to solve various research problems. In the present paper, the problem of Anomaly Network Intrusion Detection is studied. A supervised classifier based on deep learning with a conventional architecture of neural networks (CNN) is adopted to classify normal and abnormal network users behavior in order to detect attacks. Unlike usual metrics, such as the correct classification rate, used to select appropriate classifiers, the stability of the learning algorithm is used to choose an appropriate Deep Learning model for this application. Since the synaptic weights and consequently the performance of a deep learning classifier depend crucially on the optimization algorithm (gradient descent, conjugate gradient, etc), the choice of an algorithm that lead to stable performance independently of its initialization is of a great importance. A set of experiments based on the NSL-KDD dataset is achieved to compare different learning techniques in terms of accuracy and stability.