Online Grooming Detection on Social Media Platforms

Abstract

Online grooming detection has become a critical research topic in the era of extensive data analysis. It is essential to protect vulnerable users, particularly adolescents, against sexual predation on online platforms and media. However, many factors challenge online grooming detection, which leads to a high-risk problem for youth. The primary goal of this research work is to provide techniques that increase children’s security on online chat platforms. To this extent, many experiments have been conducted to create models fulfilling our research goal. As such, this thesis contains a comprehensive survey of child exploitation in chat logs that provides the readers with a deep knowledge of the problem, possible research gaps, and proposed solutions. In this research, we split the online grooming detection problem into several subproblems, including author profiling, predatory conversation detection, predatory identification, and data limitations issues. The leading theory behind the author profiling in this problem comes from the fact that online predators provide fake identities to tarp their young victims. At the same time, children’s characteristics differ from the ones who imitate a minor, which leads us to detect the gender of users in this research. In this thesis, we propose a gender detection model that can recognize the gender of authors based on their keystroke dynamics features. This research also provides a fake identity detection technique with a high performance that detects users who are dishonest about their identity. Providing an automatic predatory conversation detection system facilitates law enforcement authorities to act on time before any tragedy occurs. Therefore, we have examined and proposed several predatory conversation detection and predatory identification techniques focusing on finding the best feature vectors and embeddings that lead to the best performance in online grooming detection. This thesis also aims to gain deep knowledge about predatory behaviour with semantic analysis. We might lose some semantic information by applying conventional embeddings such as Word2vec or GloVe feature vectors since they provide a single word embedding for a term in different contexts. At the same time, humans show their motivations in phrases or sentences rather than single terms. So, we provide an online grooming detection model based on extracting embeddings from sentences rather than single words. We apply contextual model based such as Bert-based and RoBerta-based systems for each sentence. Several constraints, such as privacy and security issues, availability, and the imbalanced nature of the datasets, challenge online grooming datasets. The number of predatory chat logs is considerably lower than the other online conversations, leading to a highly imbalanced data problem. It is challenging to build a machine learning model based on imbalanced datasets, which motivates us to provide a model to handle this issue. This research proposes a model that uses a hybrid sampling and class re-distribution to gain augmented data for coping with highly imbalanced datasets. We also improve the diversity of classifiers and feature vectors by perturbing the data along with the augmentation in an iterative manner. Finally, we conclude our research by discussing potential research gaps and open problems and proposing possible solutions for them to give deep insights to the readers of future work based on the work of this thesis.