MaLang: A Decentralized Deep Learning Approach for Detecting Abusive Textual Content

Abstract

Cyberbullying is a growing and significant problem in today’s workplace. Existing automated cyberbullying detection solutions rely on machine learning and deep learning techniques. It is proven that the deep learning-based approaches produce better accuracy for text-based classification than other existing approaches. A novel decentralized deep learning approach called MaLang is developed to detect abusive textual content. MaLang is deployed at two levels in a network: (1) the System Level and (2) the Cloud Level, to tackle the usage of toxic or abusive content on any messaging application within a company’s networks. The system-level module consists of a simple deep learning model called CASE that reads the user’s messaging data and classifies them into abusive and non-abusive categories, without sending any raw or readable data to the cloud. Identified abusive messages are sent to the cloud module with a unique identifier to keep user profiles hidden. The cloud module, called KIPP, utilizes deep learning to determine the probability of a message containing different categories of toxic content, such as: ‘Toxic’, ‘Insult’, ‘Threat’, or ‘Hate Speech’. MaLang achieves a 98.2% classification accuracy that outperforms other current cyberbullying detection systems.