Towards Safer Digital Spaces: Automated Detection of Cyberbullying Through Multi-Modal Learning
M. Saravana Karthikeyan, J. Relin Francis Raj, R. Santhana Krishnan, D. Abitha Kumari, S. Murali, K. Paul Joshua
Abstract
Cyberbullying and online harassment present significant challenges to digital safety, demanding robust detection systems capable of identifying abusive content across multiple formats. This work proposes a multi-modal framework that leverages CLIP and RoBERTa embeddings to detect and categorize cyberbullying and harassment in both text and image-based posts. By utilizing two curated datasets, HarassDetect-TextImage and CyberGuard-MultiModal, the model is trained on a diverse range of offensive and non-offensive content. Text data is preprocessed through advanced techniques, while CLIP and RoBERTa generate embeddings capturing the context of both textual and visual inputs. For sequential analysis of threaded posts, a Transformer layer further enhances the model's contextual awareness, enabling it to recognize harassment patterns that unfold over multiple messages. A Support Vector Machine (SVM) classifier then categorizes content into Non-Cyberbullying, Minor, Moderate, and Severe classes, each associated with confidence levels guiding automated actions. High-confidence classifications trigger immediate responses, such as content restriction or user warnings, while moderate-confidence cases are flagged for manual review. This adaptive framework not only improves accuracy but also allows for a continuous feedback loop, updating the model to address evolving patterns of cyberbullying. Evaluation results demonstrate the efficacy of this approach in creating safer online environments through automated and contextually aware cyberbullying detection.