Hybrid Text–Image Fusion Model for Early Detection of Cyberbullying in Online Social Ecosystems

Abstract

The rapid expansion of online social ecosystems has intensified the prevalence of cyberbullying, posing significant psychological and social risks to individuals, particularly adolescents. Traditional detection systems primarily rely on textual analysis, often overlooking harmful visual content such as memes, edited images, and symbolic imagery. To address this limitation, this study proposes a Hybrid Text–Image Fusion Model for Early Detection of Cyberbullying in Online Social Ecosystems, integrating both textual and visual modalities for improved classification performance.

The proposed framework consists of two primary components: a text classification model and an image classification model. The text-based module utilizes Term Frequency–Inverse Document Frequency (TF-IDF) feature extraction combined with Logistic Regression to effectively capture linguistic patterns indicative of abusive or harmful communication. The image-based module employs a Convolutional Neural Network (CNN) architecture with convolutional, pooling, and fully connected layers to detect visual cues associated with cyberbullying. The outputs of both models are fused to generate a unified prediction, enabling multimodal analysis of social media content.

By leveraging complementary information from textual and visual inputs, the hybrid approach enhances detection accuracy and robustness compared to unimodal systems. Experimental evaluation demonstrates the model’s capability to identify early-stage cyberbullying instances with improved reliability. The proposed system contributes toward building safer online environments by enabling proactive moderation and automated monitoring across diverse social platforms.