Cross-Lingual and Multimodal Cyberbullying and Bias Detection and Content Generation via CyberGenDet
Guohao Yang
Co-Presenters: Individual Presentation
College: The Dorothy and George Hennings College of Science, Mathematics and Technology
Major: Computer Science
Faculty Research Mentor: Yulia Kumar
Abstract:
The more digital interactions occur, the greater the demand for effective mechanisms to discourage cyberbullying and bias. CyberGenDet uses OpenAI’s state-of-the-art technologies to detect and generate content in various languages and formats: text, images, and videos. This web application is unique in that, for the first time, it employs a jailbreaking technique to overcome the ethical limitations imposed on AI, enabling the creation of rich synthetic datasets that closely mimic real-world bias and cyberbullying dynamics. CyberGenDet achieves superior detection accuracy and operational flexibility by integrating multimodal AI with advanced transformer-based architectures. Moreover, its cross-lingual performance ensures efficacy across multiple linguistic and cultural settings, making it a key tool for researchers and practitioners working toward a safer online environment. In evaluations using both synthetic and real-world datasets, CyberGenDet achieved a high average detection accuracy , significantly outperforming single-modality detection systems.