metadata

language:
  - en
license: mit
tags:
  - knowledge-graph
  - rdf
  - owl
  - ontology
  - cybersecurity
annotations_creators:
  - expert-generated
pretty_name: D3FEND
size_categories:
  - 100K<n<1M
task_categories:
  - graph-ml
dataset_info:
  features:
    - name: subject
      dtype: string
    - name: predicate
      dtype: string
    - name: object
      dtype: string
  config_name: default
  splits:
    - name: train
      num_bytes: 46899451
      num_examples: 231842
  dataset_size: 46899451
viewer: false

D3FEND: A knowledge graph of cybersecurity countermeasures

Overview

D3FEND encodes a countermeasure knowledge base in the form of a knowledge graph. It meticulously organizes key concepts and relations in the cybersecurity countermeasure domain, linking each to pertinent references in the cybersecurity literature.

Use-cases

Researchers and cybersecurity enthusiasts can leverage D3FEND to:

Develop sophisticated graph-based models.
Fine-tune large language models, focusing on cybersecurity knowledge graph completion.
Explore the complexities and nuances of defensive techniques, mappings to MITRE ATT&CK, weaknesses (CWEs), and cybersecurity taxonomies.
Gain insight into ontology development and modeling in the cybersecurity domain.

Dataset construction and pre-processing

Source:

Building and Verification:

Construction: The ontology, denoted as d3fend-full.owl, was built from the beta version of the D3FEND ontology referenced above using documented README in d3fend-ontology. This includes the CWE extensions.
Import and Reasoning: Imported into Protege version 5.6.1, utilizing the Pellet reasoner plugin for logical reasoning and verification.
Coherence Check: Utilized the Debug Ontology plugin in Protege to ensure the ontology's coherence and consistency.

Exporting, Transformation, and Compression:

Note: The following steps were performed using Apache Jena's command line tools. (https://jena.apache.org/documentation/tools/)

Exporting Inferred Axioms: Post-verification, I exported inferred axioms along with asserted axioms and annotations. Detailed Process
Filtering: The materialized ontology was filtered using d3fend.rq to retain relevant triples.
Format Transformation: Subsequently transformed to Turtle and N-Triples formats for diverse usability. Note: I export in Turtle first because it is easier to read and verify. Then I convert to N-Triples.
```
arq --query=d3fend.rq --data=d3fend.owl --results=turtle > d3fend.ttl
riot --output=nt d3fend.ttl > d3fend.nt
```
Compression: Compressed the resulting ontology files using gzip.

Features

The D3FEND dataset is composed of triples representing the relationships between different cybersecurity countermeasures. Each triple is a representation of a statement about a cybersecurity concept or a relationship between concepts. The dataset includes the following features:

1. Subject (`string`)

The subject of a triple is the entity that the statement is about. In this dataset, the subject represents a cybersecurity concept or entity, such as a specific countermeasure or ATT&CK technique.

2. Predicate (`string`)

The predicate of a triple represents the property or characteristic of the subject, or the nature of the relationship between the subject and the object. For instance, it might represent a specific type of relationship like "may-be-associated-with" or "has a reference."

3. Object (`string`)

The object of a triple is the entity that is related to the subject by the predicate. It can be another cybersecurity concept, such as an ATT&CK technique, or a literal value representing a property of the subject, such as a name or a description.

Usage

First make sure you have the requirements installed:

pip install datasets
pip install rdflib

You can load the dataset using the Hugging Face Datasets library with the following Python code:

from datasets import load_dataset
dataset = load_dataset('wikipunk/d3fend', split='train')

Note on Format:

The subject, predicate, and object are stored in N3 notation, a verbose serialization for RDF. This allows users to unambiguously parse each component using rdflib.util.from_n3 from the RDFLib Python library. For example:

from rdflib.util import from_n3
subject_node = from_n3(dataset[0]['subject'])
predicate_node = from_n3(dataset[0]['predicate'])
object_node = from_n3(dataset[0]['object'])

Once loaded, each example in the dataset will be a dictionary with subject, predicate, and object keys corresponding to the features described above.

Example

Here is an example of a triple in the dataset:

Subject: "<http://d3fend.mitre.org/ontologies/d3fend.owl#T1550.002>"
Predicate: "<http://d3fend.mitre.org/ontologies/d3fend.owl#may-be-associated-with>"
Object: "<http://d3fend.mitre.org/ontologies/d3fend.owl#T1218.014>"

This triple represents the statement that the ATT&CK technique identified by T1550.002 may be associated with the ATT&CK technique identified by T1218.014.

Acknowledgements

This ontology is developed by MITRE Corporation and is licensed under the MIT license. I would like to thank the authors for their work which has opened my eyes to a new world of cybersecurity modeling.

If you are a cybersecurity expert please consider contributing to D3FEND.

D3FEND Resources

Citation

@techreport{kaloroumakis2021d3fend,
  title={Toward a Knowledge Graph of Cybersecurity Countermeasures},
  author={Kaloroumakis, Peter E. and Smith, Michael J.},
  institution={The MITRE Corporation},
  year={2021},
  url={https://d3fend.mitre.org/resources/D3FEND.pdf}
}