Enhancing Security in dynamic IoT networks through Incremental Meta-learning

The widespread adoption of Internet of Things (IoT) devices presents numerous challenges, including security vulnerabilities, scalability, data management, and reliability concerns. Device identification is of paramount importance for addressing these challenges as it serves as a foundation for implementing security measures, a monitoring system to set up access control and regulate device usage within networks. Commonly used device identifiers such as MAC and IP addresses, are vulnerable to spoofing, highlighting the need for techniques such as network traffic-based device behavior analysis. Developing a device identification system for IoT is challenging due to dynamic nature of IoT environments. Further, identifying devices is exacerbated with limited network activity traces. Previous works rely on batch learning models that require extensive labeled data and are unable to adapt to real-time data changes. In this work, we develop an incremental meta-learning framework that uses machine learning to analyze and generalize device behaviors from network traces. Integrating incremental learning algorithms into the framework allows continuous adaptation to evolving data while using meta-learner model ensures efficient model training with limited data. The proposed model outperformed recent works with an accuracy of $\mathbf{8 9 . 5 4 \%}$ and provided consistent performance across five stages of incremental training.