Sparse Bayesian Extreme Learning Machine for Intrusion Detection System by Kuok Kai Ian Thesis Supervisor: Associate Professor, Dr. Vong Chi Man In the past decades, the amount of applications and services based on Internet is rapidly growing up. Meanwhile, numerous companies are increasingly concerned with network security for various cyber attacks, hijack and data loss. For this reason, intrusion detection system (IDS) has been developed for such network security, which allows users to monitor network activities, recognize attack patterns, and detect abnormal patterns of network activities. There are two requirements for IDS detection – computational cost and data scalability. ELM is a new kind of single-hidden layer feed forward neural network with an extremely low computational costand suitable for large-scale data. However, ELM may suffer from accuracy since the solved output weights is a least squares minimization issue leading to over-fitting. As a result, the accuracy of IDS is definitely affected. An alternative technique is KELM that is a kernel-based variant of ELM. However, the relatively long execution period affects the overall performance of IDS. In this thesis, SBELM was applied in IDS data analysis in order to minimize the number of neurons in training model so that an efficient IDS detection model with high performance for IDS can be produced. In order to verify the effectiveness of SBELM, a famous benchmark dataset KDD99 was employed for experiments. From experimental results, the execution time of SBELM is only one tenth of KELM, and also 30% faster than ELM. On the other hand, SBELM is raised nearly 4% than ELM in the aspect of accuracy. SBELM outperforms KELM and ELM in accuracy and execution time, which are the two key requirements in IDS detection.