Effect of Activation Functions on the Performance of Deep Learning Algorithms for Network Intrusion Detection Systems

摘要

Increased capability and complexity of present-day networks is a product of advancements in technology which has strengthened inter-human connectivity like never before. But technological advancements empower both the developer as well as the attacker. As a result, the severity of network-based attacks have also escalated immensely. The need of the hour is to develop sophisticated intrusion detection systems that are equipped with state of the art technologies like deep learning. Several deep learning architectures for anomaly based network intrusion detection system have been proposed in literature and different authors have worked with different types of activation functions using the same algorithm and obtained different results. Due to this, performance comparison between different works based on the same algorithm differs and thus they cannot be compared. Also the use of traditional intrusion detection datasets (DARPA, KDD98, KDD99) does not provide an accurate measure of the effectiveness of deep learning algorithms for intrusion detection because these datasets lack many modem day attacks and characteristics of real time traffic. To fill these research gaps, we analyze the effect of activation functions on the performance of two deep learning algorithms: Deep Artificial Neural Network (DNN) and Convolutional Neural Network (CNN) on two recent intrusion detection datasets: NSL-KDD and UNSW-NB15 in this paper. This paper attempts to select the best activation function to tune DNN and CNN models to attain maximum accuracy in minimum time for network intrusion detection systems.