Studying the Efficiency of Deep Neural Networks as Intrusion Detection Systems

Abstract

Firewalls and antivirus software are insufficient protection against intruders in a network. Without a good detection system, a computer network can be accessed by an unauthorized individual and their malicious activities can go undetected. Such malicious users endanger the confidentiality of the data of other users in the system. Absence of a good detection system also increases the threat of a Denial of Service (DoS) attack, since the system is vulnerable to be targeted from the inside. There is thus a need for a powerful detection system that validates whether a user in the system is an authorized user or indeed a misfeasor. This can be achieved using machine learning and deep learning techniques, which were implemented in the project to build efficient and scalable models for intrusion detection. For comparison purposes, the training was done with several other classical machine learning algorithms along with DNNs for prediction and classification of attacks on a Network Intrusion Detection System (N-IDS). The dataset used for training and testing was the KDD-'99' dataset, and the evaluation was done against classical Machine Learning algorithms that are effective classifiers. The findings have been analyzed to understand if Deep Learning can be used to improve existing solutions being used in the industry. The compared results have shown that the data has an affinity for 3-layer neural networks for performance over all the machine learning algorithms.