A Hybrid Framework for Unlabeled Network Intrusion Detection: Integrating Llms With Statistical, Heuristic, and Signature-based Methods
Year: 2026
Type: article
Abstract: Computer networks have become essential infrastructures underpinning modern society, enabling personal, corporate, and industrial communications on a global scale. However, as their ubiquity increases, so too does the complexity and attack surface available to malicious actors. Consequently, ensuring the security of these networks constitutes a critical challenge, given that vulnerabilities may be exploited to compromise data integrity, disrupt services, and inflict substantial damage. Against this backdrop, this paper proposes an approach for network intrusion detection leveraging Large Language Models (LLMs). The proposed system employs LLM models such as LLaMA3 and Qwen3 to interpret unlabeled network packets, supported by statistical analyses, heuristics, YARA signatures, and entropy calculations, with the aim of identifying malicious patterns within network traffic flows. The results highlight the potential of LLMs for the semantic analysis of network traffic, while also demonstrating how complementary techniques can further enhance this capability.
Source: Cadernos Cajuína
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7236433218
