Evolving neural network intrusion detection system for MCPS

Medical Cyber Physical Systems (MCPS) are some of the most promising next generation technologies so far. Like many other systems connected to a wider network such as internet, MCPS are also vulnerable to various forms of network attacks. For detecting such diverse forms of attack, we need smart and efficient mechanisms. Human intelligence is good enough to track such attacks but when it is a huge number of traffic it is no more a feasible process to detect them manually as it is time consuming and computationally intensive. Machine learning techniques embracing artificial intelligence are emerging as powerful tools to detect abnormalities in the network data. Supervised Neural Networks are some of the most efficient techniques to perform such classification. In this paper, we propose an evolving neural network technique that evolves based on classification, elimination and prioritization while focusing on time, space and accuracy to efficiently classify the four major types of network attack traffic found in an effectively pruned KDD dataset. We also show a leap of performance with hyper-parameter optimization which highly enhances the benefit of our proposed mechanism. Finally, the new performance gain is compared with a boosted Decision Tree. We believe our proposed mechanism can be adopted to new forms of attack categories and sub-categories.