Abstract:
An automated real-time detection is required for preventing cyber bullying and abusive be-havior in cyber spaces such as social media, rather than penalizing afterward. Researchers till now have focused on mostly textual data analysis for the purpose of detection that can become ine ective in case of experiencing varieties of languages, code mixing, code switching, usage of abusive words in normal conversations without having any intentions of bullying, etc. To this extent, it becomes utmost important to devise a generalized mechanism to detect cyber bullying going beyond the mentioned limitations. To do so, in this paper, we investigate de-tection of cyber bullying through analyzing facial expressions that generally o er a generalized metric irrespective of diversity in languages and their usages. Here, our focus is on nding common patterns using the phenomenon of facial expression alteration while cyber bullying. Consequently, we propose a real-time cyber bullying detection mechanism, which can achieve an accuracy up to 100% and an average of 98% while trained with individual's own facial dynamics and an accuracy up to 74% and an average of 67% while trained with other persons' facial dynamics using a machine learning algorithm. We con rm these ndings through a real implementation of our proposed mechanism and experimentation over several users using the implementation. Thereafter, we achieve up to 76% accuracy by experimenting with synthe-sized data, which agrees with our real data experimentation. Experimental results demonstrate that four facial dynamics have the potential to become key factors in the detection of cyber bullying.
