Predicting malware threat intelligence using KGs

Nidhi Rastogi, Sharmishtha Dutta, Ryan Christian, Jared Gridley, Mohammad Zaki, Alex Gittens, Charų C. Aggarwal

Abstract

Cybersecurity analysts face an overwhelming volume of unstructured threat intelligence. This paper introduces TINKER, an end-to-end framework that transforms disparate threat reports into a structured Knowledge Graph (CTI-KG). By employing semi-supervised information extraction and graph embeddings, TINKER enables the prediction of missing threat intelligence. We present two datasets: CS3K (manually curated) and CS40K (automatically generated with 40,000 triples). Our evaluation using the TuckER model achieves a Hits@10 of 80.4 and a Mean Reciprocal Rank (MRR) of 0.75, demonstrating the system's ability to accurately infer hidden relationships in malware data.

Introduction

The rapid proliferation of cyber threats has led to a surge in Cyber Threat Intelligence (CTI) published in unstructured formats like blog posts and whitepapers. While rich in detail, this data is difficult for automated systems to ingest and often lacks the semantic context needed to understand complex attack patterns. Furthermore, threat information is frequently fragmented, leaving gaps in our understanding of malware authors, targets, and techniques. TINKER addresses these challenges by automatically extracting entities and relations into a Malware Knowledge Graph (MalKG), facilitating both structured search and predictive inference.

Methodology: The TINKER Framework

The TINKER pipeline automates the transition from raw text to actionable intelligence through four main stages:

  1. Data Collection: Aggregating over 1,100 threat reports from leading security vendors.
  2. Information Extraction (IE): Utilizing a semi-supervised approach for Named Entity Recognition (NER) and Relation Extraction (RE). This identifies key entities (e.g., malware families, vulnerabilities) and the semantic links between them (e.g., "drops," "targets").
  3. Ontological Mapping: Aligning extracted data with established cybersecurity ontologies like MALOnt to ensure logical consistency.
  4. Knowledge Graph Completion: Applying the TuckER model—a tensor decomposition-based embedding method—to predict missing links within the graph.
The TINKER framework pipeline from unstructured reports to predictive intelligence.

Datasets

The research introduces two benchmark datasets for malware threat intelligence. As shown in the dataset statistics, CS3K serves as a manually annotated gold standard, while CS40K provides the scale necessary for training robust embedding models. CS40K contains approximately 40,000 triples generated from over 27,000 unique entities.

Results and Discussion

The core strength of TINKER lies in its predictive capabilities. By treating threat intelligence as a graph completion problem, the system can infer missing attributes of a threat actor or malware sample.

Experimental results show that the TuckER model significantly outperforms traditional translation-based models. On the large-scale CS40K dataset, the model achieved a Hits@10 score of 80.4, indicating that the correct prediction appears in the top 10 results in over 80% of cases.

Performance metrics of the TuckER model on the CS40K dataset.

Case studies further validate the system's utility. For instance, TINKER successfully identified specific malware families and predicted regional targets for themed attacks even when those links were not explicitly stated in the source text.

Conclusion

TINKER demonstrates that Knowledge Graphs are a powerful tool for managing and predicting cyber threat intelligence. By moving beyond simple keyword matching to semantic understanding and link prediction, the framework provides a foundation for more proactive and automated defense strategies. Future work will focus on handling the dynamic nature of evolving threats and integrating the CTI-KG directly into real-time security orchestration platforms.

Summary updated

Abstract

Large amounts of threat intelligence information about malware attacks are available in disparate, typically unstructured, formats. Knowledge graphs can capture this information and its context using RDF triples represented by entities and relations. Sparse or inaccurate threat information, however, leads to challenges such as incomplete or erroneous triples. Generic information extraction (IE) models used to populate the knowledge graph cannot fully guarantee domain-specific context. This paper proposes a system to generate a Malware Knowledge Graph called MalKG, the first open-source automated knowledge graph for malware threat intelligence. MalKG dataset (MT40K\footnote{ Anonymous GitHub link: this https URL}) contains approximately 40,000 triples generated from 27,354 unique entities and 34 relations. For ground truth, we manually curate a knowledge graph called MT3K, with 3,027 triples generated from 5,741 unique entities and 22 relations. We demonstrate the intelligence prediction of MalKG using two use cases. Predicting malware threat information using the benchmark model achieves 80.4 for the hits@10 metric (predicts the top 10 options for an information class), and 0.75 for the MRR (mean reciprocal rank). We also propose an automated, contextual framework for information extraction, both manually and automatically, at the sentence level from 1,100 malware threat reports and from the common vulnerabilities and exposures (CVE) database.
BibTeX
@inproceedings{rastogi2021predicting,
  author = {Nidhi Rastogi and Sharmishtha Dutta and Ryan Christian and Jared Gridley and Mohammad Zaki and Alex Gittens and Charų C. Aggarwal},
  title = {{Predicting malware threat intelligence using KGs}},
  booktitle = {TrustCom},
  year = {2021},
  doi = {10.1109/trustcom56396.2022.00225},
}
Figure