Deep transfer learning-based network traffic classification for scarce dataset in 5G IoT systems

Abstract

Internet of Things (IoT) can provide the interconnection and data sharing among devices, vehicles, buildings via various sensors with the development of 5G, and it has been widely used in different services such as e-commerce, heath-care, smart buildings. In the meantime, various cyber-attacks for IoT have increased and caused huge losses. Lots of security mechanisms are rapidly being proposed to prevent the potentially malicious attackers for IoT, in which machine learning especially deep learning (DL) as increasingly popular solution for security has been implemented in intrusion detection system (IDS) and others. However, the lack of enough datasets prevents the application of IDS in 5G IoT system. As one of fundamental components of IDS, network traffic classification shows a discretization, individualization and fine-grained trend which derives the different personalized classification methods for different requirements and scenarios. In this case, the data-driven DL faces the following challenges. First, there are only a few labeled datasets in the various personalized application scenarios, which undoubtedly limits the deployment of DL classification. Second, not all scenarios have rich computing capability for that training a neural network requires lots of computing resources. Therefore, this paper proposes a traffic classification method based on deep transfer learning for 5G IoT scenarios with scarce labeled data and limited computing capability, and trains the classification model by weight transferring and neural network fine-tuning. Different from the previous work that extract artificially designed features, the proposed method retains the end-to-end learning performance of DL and reduces the risk of suffering concept drift to reduce human intervention. Experimental results show that when only 10% of dataset are used to label the data samples, the classification accuracy is close to the results of full training dataset.