POSTER : Detection of Online Radical Content Using Multimodal Approach

Several criminal and terrorist organisations have benefited tremendously from the worldwide reach, growth, and speed of the Internet. By utilising the Internet and multiple social media platforms, they are now able to spread their views, widen their reach, and have opportunities to recruit people from all over the world. This has also given them a media platform to broadcast their messages and different propaganda material aiming to spread fear, radicalise and recruit potential members. Previous study has demonstrated that the use of Internet by terrorist groups has significantly increased in the recent years [1]. Several social media platforms such as Twitter and Facebook are working towards keeping these platforms clean by suspending those who are promoting violent content or extremist behaviour. However, due to the volume and speed of the generated data it is still challenging to detect those misbehaving users accurately and in a timely manner. Recent research has focused on studying the online behaviour of pro-extremists users mainly by performing contentbased analysis in order to identify distinguishing textual features that can aid in automatic detection of these users [2]. However, using this approach alone has several shortcomings including producing a large number of false positives, having a high dependency on the data, and it can be easily evaded by adapting the users writing styles through automated tools. Ashcroft et al. make an attempt to automatically detect Jihadist messages on Twitter [2]. They adopt a machine-learning method to classify tweets as ISIS supporters or not. They focus on English tweets that contain a reference to a set of predefined English hashtags related to ISIS. One of the limitations of their approach is that it is highly dependant on the data. Choudhary et al., [3] surveyed existing literature on counter terrorism and social network analysis. Some of the studied problems in this domain are related to identifying key-players, finding behaviour patterns, community discovery, and disrupting terrorist networks. They found that the use of Social Network Analysis (SNA) is one of the most successful methods for counter terrorism in social networks. Building on the findings of previous research efforts, in this paper we propose a novel method to detect online extremist content that is based on multi-modal approach including textual (syntactic and semantic) features, behavioural features based on social network analysis, as well as psychologicalbased features. We study the effects of adding these psychological, and personality features to the accuracy of our model using Linguistic Inquiry and Word Count (LIWC). We perform an experiment on the Twitter platform using our approach with the aim of detecting radical content and proextremist tweets. We adopt machine learning methodology to classify tweets and use our proposed approach for features identification. We envisage that this approach can be utilised by law enforcement investigators and security analysts to aid in detecting and limiting online radical propaganda.