Malicious data classification in packet data network through hybrid meta deep learning

Abstract

Advancements in wireless network technology have provided a powerful tool to boost productivity and serve a strong communication which overcomes the limitations of wired networks. However, because of using wireless networks, security is an increasing concern among the community. At the time of our study, we are in the era of 5G networks. Although we are in the 5th generation of telecommunication we are still struggling with security. The upcoming generation, 6G, aims to solve the security concerns by providing a secure and trust networking system. In our study, we aim to integrate AI and more advanced infrastructure which will provide a tremendous solution in this regard. In order to deal with this issue we primarily aim to come up with a solution that provides a reliable intrusion detection system in spite of being trained with a small amount of data. In our study, we aim to integrate AI and more advanced infrastructure which will provide a tremendous solution in this regard. Thus, we employed a trusted networking system based on AI. Here, at first we primarily focused on Reinforcement Learning (RL) to classify the network data coming from the untrusted packet data networks (PDN), whether it is malicious or not. Another existing problem is people currently rely on machine learning techniques to create a trustworthy networking system. However, it hinders the development of getting a reliable network as the number of real publicly available malicious data is not sufficient to train a model properly and in real life people are not very keen to share these data as they are sensitive. Therefore, we propose a novel idea of hybrid meta learning in the detection of malicious packet data. We use a combination of Siamese and Prototypical network where Siamese network is used for binary classification and Prototypical network is used for multi class classification. As both approaches are based on meta learning techniques, it requires a very small amount of data. By utilizing this characteristic of meta learning, we were able to train our model with just 3000 data samples and achieve more than 90% accuracy for both meta learning tactics. Lastly we provide a comprehensive study on the given RL methods and hybrid meta learning and share our future thoughts. The purpose of our study is to provide a secure and trustworthy network domain which enhances the communication between end users.