AI Powered Multimodal Content Moderation for Online Safety in Social Media Platforms
Charan Kumar V, C L L Gowda, Ganesana Charishma, Darshan B.M, Lakshmi Prasad D Yadav, Vidya Vidya · 2025
In an era where social media platforms are inundated with diverse and complex content, moderation is the answer to sieving out toxic content. Platforms such as Twitter, Facebook, and Instagram enable global connectivity, but also have spaces for toxic content in the way of fake news, hate speech, cyberbullying, and deepfakes that impede mental health and generate economic losses. Sifting and screening such content is an urgent need. This paper presents a new approach to content moderation based on a Multimodal Social Media Content Moderation framework established by Hybrid Graph Theory and Bio-inspired Optimization (MSCMGTB). The model utilizes Convolutional Neural Networks (CNNs) for extracting visual features and Transformer-based models for text understanding. Multimodal inputs are dynamically aligned by a Bi-directional Attention Mechanism (BAM), whereas hyperparameters are optimized using Genetic Algorithms (GA). The model outperforms existing approaches, such as SGNN and CrediBot, with improved precision, accuracy, recall, AUC, specificity, and reduced response delays. In addition, the research explores image captioning as a content moderation task with a human-machine collaboration strategy. With training on Flickr30k and MS Coco, the model reduces review times by 13 %, where finetuning reduces it by 28 %. This recognizes the power of humanmachine collaboration in preventing review backlogs. The model also yields automatic region-based obfuscation with a top1 accuracy of 90.3 annotation. Simulations of real-world use demonstrate improved consistency of decisions, reduced variance of review time, and scalability in dynamic social media environments, exhibiting MSCMGTB's efficacy and usability for real-time application.