Cyber threat intelligence depends on the timely identification of emerging threats, and social media platforms like X are a valuable source of early signals. However, manually classifying the flood of security-related tweets is labor-intensive and slow, creating a bottleneck for analysts who need situational awareness.
The paper proposes a teacher-student framework in which a large language model acts as a teacher, producing weak annotations for unlabeled tweets. A student model then learns from these noisy labels, offering a way to build classifiers without extensive human-annotated datasets. This approach leverages the LLM's language understanding to capture threat-relevant context.
The significance lies in its potential to automate the triage of cyber threat tweets, helping organizations and analysts keep pace with fast-moving risks. By reducing the manual effort required for labeling, the framework could make CTI pipelines more scalable and responsive to new threats as they emerge on social media.