Peersman, Claudia (2018) Detecting deceptive behaviour in the wild:text mining for online child protection in the presence of noisy and adversarial social media communications. PhD thesis, UNSPECIFIED.
2018peersmanphd.pdf - Published Version
Available under License Creative Commons Attribution-NoDerivs.
Download (2MB)
Abstract
A real-life application of text mining research “in the wild”, i.e. in online social media, differs from more general applications in that its defining characteristics are both domain and process dependent. This gives rise to a number of challenges of which contemporary research has only scratched the surface. More specifically, a text mining approach applied in the wild typically has no control over the dataset size. Hence, the system has to be robust towards limited data availability, a variable number of samples across users and a highly skewed dataset. Additionally, the quality of the data cannot be guaranteed. As a result, the approach needs to be tolerant to a certain degree of linguistic noise. Finally, it has to be robust towards deceptive behaviour or adversaries. This thesis examines the viability of a text mining approach for supporting cybercrime investigations pertaining to online child protection. The main contributions of this dissertation are as follows. A systematic study of different aspects of methodological design of a state-ofthe- art text mining approach is presented to assess its scalability towards a large, imbalanced and linguistically noisy social media dataset. In this framework, three key automatic text categorisation tasks are examined, namely the feasibility to (i) identify a social network user’s age group and gender based on textual information found in only one single message; (ii) aggregate predictions on the message level to the user level without neglecting potential clues of deception and detect false user profiles on social networks and (iii) identify child sexual abuse media among thousands of legal other media, including adult pornography, based on their filename. Finally, a novel approach is presented that combines age group predictions with advanced text clustering techniques and unsupervised learning to identify online child sex offenders’ grooming behaviour. The methodology presented in this thesis was extensively discussed with law enforcement to assess its forensic readiness. Additionally, each component was evaluated on actual child sex offender data. Despite the challenging characteristics of these text types, the results show high degrees of accuracy for false profile detection, identifying grooming behaviour and child sexual abuse media identification.