Cyber Social Threats

Policymakers and practitioners assert serious concerns, due to potentially unfair and untraceable socio-cultural implications of AI, notably in the cyber social domain. Recognition of harmful narratives is challenging due to their linguistic, lexical, and contextual characteristics. My research explores the problems of online extremism, toxicity, and disinformation to examine the behavioral change in these persuasive discourses.

In my work on analyzing Islamist extremist narratives, a key challenge is disambiguating content because the meaning of the same lexical item may change (e.g., jihad) in different contexts. This work distinguished a group of outlier (not extremist) users from the ground truth extremist user dataset, while obtaining a significant improvement in performance. This finding revealed a potential unfair social discrimination at scale. I have also been investigating online toxic behavior revealing the effect of context and multimodality in the interactions between aggressor and victim, designing a new method to learn coherent representations interactional data.