Using deep learning to detect social media ‘trolls’

Detecting criminal activity online is not a new concept but how it can occur is changing. Technology and the influx of social media applications and platforms has a vital part to play in this changing landscape. As such, we observe an increasing problem with cyber abuse and ‘trolling’/toxicity amongst social media platforms sharing stories, posts, memes sharing content. In this paper we present our work into the application of deep learning techniques for the detection of ‘trolls’ and toxic content shared on social media platforms. We propose a machine learning solution for the detection of toxic images based on embedded text content. The project utilizes GloVe word embeddings for data augmentation for improved prediction capabilities. Our methodology details the implementation of Long Short-term memory Gated recurrent unit models and their Bidirectional variants, comparing our approach to related works, and highlighting evident improvements. Our experiments revealed that the best performing model, Bidirectional LSTM, achieved 0.92 testing accuracy and 0.88 inference accuracy with 0.92 and 0.88 F1-score accordingly.