Intrusion Detection System Based on Deep Learning

Abstract

Intrusion Detection Systems (IDSs) have a significant role in all networks and information systems in the world to earn the required security guarantee. IDS is one of the solutions used to reduce malicious attacks. As attackers always changing their techniques of attack and finding alternative attack methods, IDSs must also evolve in response by adopting more sophisticated methods of detection. The huge growth in the data and the significant advances in computer hardware technologies resulted in the emergence of new studies in the field of deep learning, including intrusion detection. Deep learning is a sub-field of Machine Learning (ML) methods that are based on learning data representations. Because deep learning has the potential to extract better representations from the data to create much better models, this work proposes an IDS based on deep learning techniques using Recurrent Neural Network (RNN) algorithm. Hence, this thesis presents the design and implementation of the binary class IDS based on recurrent neural networks. The performance of this system is studied with different architectures such as the number of hidden layers, number of hidden units in each layer and other parameters that impact the accuracy. This is necessary to choose a suitable RNN architecture that deals with the intrusion detection problem. All of these are applied with the benchmark KDD-99 dataset. Next, the implementation of the Crow Swarm Optimization (CSO) algorithm to reduce the dataset features has been done because reducing the features means dealing with less data, which reflects positively on the accuracy of the system and the implementation time. The fitness of this CSO algorithm has been calculated by the RNN, So the RNN algorithm used as a fitness evaluator this gives CSO algorithm more strength because actual training done on all selected features by the CSO algorithm. 5 The experimental results have illustrated that RNN is very suitable for solving the intrusion detection problem due to the high-performance results of the RNN algorithm applied to the KDD-99 dataset in binary classification methods. Finally, the obtained results for application of the CSO algorithm for feature selection have shown the superiority of the algorithm over features selection and reduction, where it produced three selected features with an accuracy rate of (98.34%), which reflects another success for the performance of our proposed DL based IDS.