An IoT intrusion detection framework based on feature selection and large language models fine-tuning.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Network Security

Background:

• Internet of Things (IoT) devices increase network vulnerabilities.
• Existing intrusion detection systems (IDS) struggle with feature redundancy and class imbalance.
• Current methods often address these challenges in isolation.

Purpose of the Study:

• To propose a novel Feature Selection and Large Language Models (LLMs)-based IoT intrusion detection framework (FSLLM).
• To address both feature redundancy and class imbalance in IoT intrusion detection.
• To enhance the efficiency and accuracy of IoT security systems.

Main Methods:

• A multi-stage feature selection algorithm combining Minimum Redundancy Maximum Relevance (mRMR) and a Pearson Correlation Coefficient (PCC)-improved Covariance Matrix Adaptation Evolution Strategy (CMA-ES).
• Fine-tuning Large Language Models (LLMs) with selected features to generate synthetic attack samples.
• Utilizing a Focal Loss (FL) function-improved LightGBM classifier for improved detection.

Main Results:

• The FSLLM framework significantly reduces redundant features by over 80%.
• Achieves comparable or superior accuracy to state-of-the-art methods across five benchmark IoT datasets.
• Demonstrates effective generation of higher-quality attack samples, mitigating class imbalance.

Conclusions:

• The proposed FSLLM framework offers an effective solution for IoT intrusion detection.
• It successfully tackles feature redundancy and class imbalance, leading to enhanced security.
• FSLLM presents a computationally efficient and accurate approach to bolstering IoT network defenses.