Answers are grounded in this paper and cite their sources.
Open-source intelligence is extensively utilized by enterprises and organizations to build intelligence networks for mitigating cyber threats. However, industries such as healthcare and finance often handle highly sensitive personal information, posing significant challenges for privacy-preserving intelligence sharing. A commonly adopted privacy protection approach is data de-identification, which involves various de-identification techniques; however, cross-referencing datasets can still result in the re-identification of original information. To address this challenge, this study proposes a novel framework leveraging LLaMA-3, an open-source generative model, fine-tuned using Low-Rank Adaptation (LoRA) for threat intelligence de-identification. This approach facilitates secure data sharing while preserving privacy. Furthermore, 8,435 real-world threat intelligence records were collected from a threat intelligence platform and categorized across multiple industries. To evaluate the effectiveness of the proposed de-identification technique, a classification model was trained and tested for industry classification. The results indicate that the industry classification model achieved an F1 score of 99% on the original dataset and 86% on the de-identified dataset. In contrast, a baseline method based on categorical replacement technology, which employs a Named Entity Recognition (NER) model to detect entities within the text and substitutes them directly with predefined categorical labels, achieved only a 75% F1 score. These findings highlight the superiority of the fine-tuned generative model, which generates context-aware and semantically coherent substitutions that more effectively preserve both the linguistic and analytical integrity of threat intelligence compared to categorical replacement. This capability enhances privacy-preserving intelligence sharing while maintaining the analytical utility and industry-specific relevance of the data. • This study proposes fine-tuned LLaMA-3 with LoRA for secure, private intelligence sharing. • The study achieved an 86% F1-Score, preserving industry traits—12% better than prior methods. • A BERT-based classification model validates authenticity, ensuring credibility of shared data.
No takes yet. Share an insight, caveat, or question.
Sun et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: