Multi-Class Classification for IoT Attack Detection: An LLM-Based Approach

Abstract

With the widespread adoption of IoT devices, cyberattacks targeting IoT networks have increased, making the detection of security threats a significant research topic. Traditional ML-based approaches exist for mitigating these threats; however, they suffer from certain limitations, such as limited generalization ability and practical applicability. In this study, a LLM-based approach is proposed for intrusion detection in IoT networks. The open-source CIC IoT 2023 dataset was utilized for attack detection in IoT networks. The detection process was conducted using the DistilBERT model. Initially, the contribution of all features in the dataset to the model’s classification performance was examined. Subsequently, the most important 30 features were selected using the RFE method, and the model’s performance was analyzed. Before feature selection, the fine-tuned DistilBERT model achieved an accuracy of 81%, whereas, after feature selection and further fine-tuning, the model’s accuracy improved to 94%. The obtained results demonstrate that feature selection significantly enhances model performance.