Document Type

Theses, Ph.D

Rights

Available under a Creative Commons Attribution Non-Commercial Share Alike 4.0 International Licence

Disciplines

1.2 COMPUTER AND INFORMATION SCIENCE, Computer Sciences

Publication Details

Thesis successfully submitted to the Technological University Dublin for the degree of Doctor of Philosophy.

Abstract

Conceptually, a vulnerability is: "A flaw or weakness in a system’s design, implementation,or operation and management that could be exploited to violate the system’s security policy".Some of these flaws can go undetected and exploited for long periods of time after soft-ware release. Although some software providers are making efforts to avoid this situ-ation, inevitability, users are still exposed to vulnerabilities that allow criminal hackersto take advantage. These vulnerabilities are constantly discussed in specialised forumson social media. Therefore, from a cyber security standpoint, the information found inthese places can be used for countermeasures actions against malicious exploitation ofsoftware. However, manual inspection of the vast quantity of shared content in socialmedia is impractical. For this reason, in this thesis, we analyse the real applicability ofsupervised classification models to automatically detect software vulnerability com-munication in expert social media channels. We cover the following three principal aspects: Firstly, we investigate the applicability of classification models in a range of 5 differ-ent datasets collected from 3 Internet Domains: Dark Web, Deep Web and SurfaceWeb. Since supervised models require labelled data, we have provided a systematiclabelling process using multiple annotators to guarantee accurate labels to carry outexperiments. Using these datasets, we have investigated the classification models withdifferent combinations of learning-based algorithms and traditional features represen-tation. Also, by oversampling the positive instances, we have achieved an increaseof 5% in Positive Recall (on average) in these models. On top of that, we have appiiplied Feature Reduction, Feature Extraction and Feature Selection techniques, whichprovided a reduction on the dimensionality of these models without damaging the accuracy, thus, providing computationally efficient models. Furthermore, in addition to traditional features representation, we have investigated the performance of robust language models, such as Word Embedding (WEMB) andSentence Embedding (SEMB) on the accuracy of classification models. RegardingWEMB, our experiment has shown that this model trained with a small security-vocabulary dataset provides comparable results with WEMB trained in a very large general-vocabulary dataset. Regarding SEMB model, our experiment has shown thatits use overcomes WEMB model in detecting vulnerability communication, recording 8% of Avg. Class Accuracy and 74% of Positive Recall. In addition, we investigate twoDeep Learning algorithms as classifiers, text CNN (Convolutional Neural Network)and RNN (Recurrent Neural Network)-based algorithms, which have improved ourmodel, resulting in the best overall performance for our task.

DOI

https://doi.org/10.21427/7fnh-2871

Download


Share

COinS