Abstract
Security analysts prepare threat analysis upon investigating an attack, an emerging cyber threat, or a recently discovered vulnerability. Threat intelligence on malware attacks and campaigns is shared on blog posts, reports, analyses, and tweets with varying technical details. Other security analysts use this intelligence to inform them of emerging threats, indicators of compromise, attack methods, and preventative measures. Collectively known as threat intelligence, it is typically in an unstructured format and, therefore, challenging to integrate seamlessly into existing IDPS systems. In this paper, we propose a framework that aggregates and combines CTI - the openly available cyber threat intelligence information. The information is extracted and stored in a structured format using knowledge graphs such that the semantics of the threat intelligence can be preserved and shared at scale with other security analysts. We propose the first semi-supervised open-source knowledge graph (KG) framework, TINKER, to capture cyber threat information and its context. Following TINKER, we generate a Cyberthreat Intelligence Knowledge Graph (CTI-KG). We demonstrate the efficacy of CTI-KG using different use cases and its application for security analysts.
Introduction
Cyber Threat Intelligence (CTI) is critical for modern defense, yet it faces a significant delivery challenge. Most intelligence is disseminated via unstructured formats like blog posts, PDFs, and social media. While rich in context, these sources are difficult to integrate into automated defense systems such as Intrusion Detection Systems (IDS).
Existing standards like STIX or CVE provide lists of indicators but often lack the semantic and contextual associations—the "why" and "how"—that connect a specific piece of malware to a vulnerability, a threat actor, and a geographic target. Security analysts are currently forced to manually "connect the dots" across thousands of disparate reports. TINKER aims to automate this by building a reasoning engine for cyber threats.
The TINKER Framework
The TINKER framework utilizes a multi-stage pipeline to transform raw text into a predictive knowledge graph (CTI-KG). The process begins with automated data collection, scraping over 1,100 threat advisory reports from major security firms and repositories spanning 2010–2022.
Information extraction is performed using Named Entity Recognition (NER) via the Flair library and entity expansion through SetExpan. Relationships are identified using a distantly supervised neural network. To ensure logical consistency, entities are mapped to established cyber-ontologies like MALOnt. The core innovation lies in Knowledge Graph Completion using the TuckER model, a linear model based on tensor decomposition. This allows the system to treat the knowledge graph as a 3D matrix and mathematically predict missing links (triples) that were not explicitly stated in the source text.
Evaluation and Results
The framework was validated using two primary datasets: CS3K, a gold-standard dataset of 85 expert-annotated reports (~3,000 triples), and CS40K, a large-scale semi-supervised dataset containing 40,000 triples and over 27,000 unique entities.
As shown in the results for the CS40K dataset, the TuckER model achieved a Hits@10 of 80.4%, meaning the correct missing piece of intelligence was in the model's top 10 suggestions over 80% of the time. Furthermore, the Hits@1 reached 73.9%, indicating that the model's primary prediction was correct nearly three-quarters of the time. Real-world case studies further demonstrated this efficacy; for instance, the model correctly identified the "Stealer" malware family associated with a specific URL indicator and inferred geographic targets for phishing campaigns that were not explicitly linked in the training data.
Discussion and Conclusion
TINKER demonstrates the power of link prediction in CTI, showing that global patterns across thousands of attacks can be used to infer missing details in individual reports. By focusing on the behavioral story—linking the "who," "how," and "where"—rather than just simple indicators, the framework provides a more comprehensive defense context.
While highly effective, the system faces challenges such as data noise from conflicting reports and a continued reliance on human experts to refine results and eliminate false positives. Future work will focus on expanding the corpus to include real-time data sources like social media to improve the "freshness" and coverage of the CTI-KG.
Summary updated