Word Representation for Cyber Security Vulnerability Domain

This repo provides a word representation (SecVuln_WE) and a dataset for benchmarking word similarity and relatedness for cyber security vulnerability domain. The following paper describes the step-by-step procedure for training the word embedding and construction of similarity dataset.

SecVuln_WE

A word2vec model trained on multiple heterogeneous sources including Vulners, English Wikipedia (Security category), Information Security Stack Exchange Q&As, Common Weakness Enumeration (CWE) and Stack Overﬂow.

Model file

The pre-trained WE (SecVuln) is stored in a .bin file (of approximate size 160 MB).

Instructions on how to use the model

Prerequisites

To load the model you will need Python 3.5 and the gensim library.

Loading the model

from gensim.models.keyedvectors import KeyedVectors
word_vect = KeyedVectors.load_word2vec_format("vulner_embedding.bin", binary=True)

Querying the model

Examples of semantic similarity queries

words=['vulnerability','patch']
for w in words:
    try:
        print(word_vect.most_similar(w)[:5])
    except KeyError as e:
            print(e)
>> [(u'vulnerabilities', 0.889), (u'bug', 0.786), (u'flaw', 0.742), (u'exploit', 0.740), (u'issues', 0.739)]
>> [(u'patches', 0.816), (u'updates', 0.707), (u'fixes', 0.702), (u'fix', 0.688), (u'upgrade', 0.667)]

print(word_vect.similarity('bug', 'flaw'))
>> 0.72691536

print(word_vect.doesnt_match("exploit attack weakness python".split()))
>> python

Examples of analogy queries

print(word_vect.most_similar(positive=['exploit', 'title'], negative=['ubuntu']))
>> [(u'vulnerability', 0.571), (u'xss', 0.556), (u'injection', 0.501)]

Word Similarity Dataset

Word Similarity dataset is a collection of words for measuring the similarity and relatedness of cyber security words. The dataset file is available here for download. The file is in csv format and consists of two columns with word1 and word2. The dataset is available here for download.

unsw-cse-soc / Vul_Word2Vec

readme