Cyberbullying detection through image classification: D-Net architecture

Social media platforms are becoming a stage where it can badly affect user&s;s mental health and well-being. Text analysis performed by many research studies to find cyberbullying but visual, graphics and images-based studies are limited. This study modified Cyberbullying Visual Indicators Dataset or CVID dataset using data augmentation to improve the outcome. It has 668 original images marked into 4 types of abuse, curse, discourage, threat and with help of semantic data augmentation total number of images increased to 1,924. A classification model was built using pre-trained DenseNet201 architecture and some changes were made like adding flattening layer, dropout with rate 0.6 and batch normalization layers to improve the model&s;s accuracy. The proposed model D-Net gave best result with 99% accuracy and showed stable performance in five-fold cross-validation, performing better than popular models such as VGG16, VGG19, MobileNetV2, Inception-V3 and Vision Transformer. This research shows strong proof that deep learning can detect many types of visual cyberbullying. In future, CVID will be improved by adding images with different lighting, making datasets bigger and more diverse, and combining it with hazard-analysis framework to create stronger and platform-independent cyberbullying prevention tools.