AeroSec: a vertical domain large language model for air traffic management cybersecurity

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Aerospace Engineering

Background:

• The increasing complexity of smart civil aviation presents critical cybersecurity challenges for air traffic management (ATM).
• Existing large language models (LLMs) in ATM do not address cybersecurity needs, creating a gap in security and safety.
• Advanced protection methods are essential due to the expanding attack surface in ATM systems.

Purpose of the Study:

• To develop a specialized large language model (LLM) for air traffic management (ATM) cybersecurity.
• To address the neglect of cybersecurity in current ATM-focused LLMs.
• To enhance the security and safety transition in civil aviation.

Main Methods:

• A prompt-engineering approach was used to generate question-answer pairs from multi-source data for ATM cybersecurity.
• A specialized fine-tuning dataset and benchmark for ATM cybersecurity were constructed.
• Two models (Deepseek-llm-7B-base and DeepSeek-R1-Distill-Qwen-14B) were fine-tuned using instruction and reasoning methods with LoRA and full-parameter techniques to create 'AeroSec'.

Main Results:

• The constructed ATM cybersecurity fine-tuning dataset and benchmark facilitated the development of the 'AeroSec' LLM.
• Fine-tuning demonstrated the effectiveness of instruction and reasoning methods combined with LoRA and full-parameter techniques.
• Validation against base and third-party models confirmed 'AeroSec's' capability in specialized ATM cybersecurity domains.

Conclusions:

• 'AeroSec' represents a significant advancement in applying LLMs to ATM cybersecurity.
• The developed method and dataset provide a foundation for future research in vertical domain LLMs for specialized cybersecurity.
• The study validates the utility of 'AeroSec' for analyzing ATM network threats and offensive-defensive strategies.