Cyberattack event and arguments extraction based on feature interaction and few-shot learning

Cyber Threat Intelligence (CTI) serves as a crucial indicator of cybersecurity events, providing critical evidence for analyzing cyberattacks and vulnerabilities. While existing deep learning approaches have shown value in processing urgent CTI data, accurately extracting cybersecurity event types and their associated arguments remains challenging due to CTI's inherent complexity and data scarcity. Moreover, the lack of a large-scale annotated corpus in the cybersecurity domain prevents direct application of large language models (LLMs) to surpass the current methods' performance. To enhance CTI information extraction efficiency, we propose the Cyber Attack Feature Interaction Information Extraction (CAFIIE) method. Our investigation reveals prevalent latent deep features in CTI threat entity interactions. Consequently, we implement multi-layer shallow feature interactions to capture these latent characteristics of threat entities and their relations. The CAFIIE model is further fine-tuned through few-shot learning to improve CTI feature utilization efficiency, making it particularly suitable for data-limited cybersecurity scenarios. Through experiments on three public and one private datasets, the CAFIIE method demonstrates superior accuracy and F1-score performance compared to existing baseline methods.