Large Language Model and Variational Autoencoder Based Deep Neural Framework for Cyber Attack Detection

Zhi Liu, Jyotheesh Gaddam, Shuo Huang, Ishara Bandara, Ming Liu, Sutharshan Rajasegarar, Muneeb Ul Hassan, Luxing Yang, Gang Li, Maia Angelova

Abstract: Recent advancements in large language models (LLMs) have enabled the development of diverse applications through efficient fine-tuning and prompt engineering. With the increasing frequency of cyberattacks targeting computer networks and connected devices, the accurate and timely detection of emerging cyber threats has become more critical than ever. This paper proposes a deep neural framework that integrates large language model and variational autoencoder (VAE) for anomaly detection, that aid in detecting emerging cyberattacks. Further, the proposed framework explores three fusion mechanisms, combining the generic representations of LLMs with the data-specific representations of VAEs, and utilising a one-class support vector machine for anomaly detection. Evaluation on benchmark datasets and ablation study demonstrates the superiority of the proposed approach in improving the anomaly detection performance by achieving F1-scores greater than 0.95 across all benchmark datasets.

External IDs:doi:10.1007/978-981-96-8183-9_8