An efficient automated multi-modal cyberbullying detection using decision fusion classifier on social media platforms

Cyberbullying poses a dangerous threat to teens, children, young people, and adults in today’s digital world. Bullies use virtual social media platforms like Twitter, Facebook, YouTube, and Instagram to harm their victims. Therefore, this online bullying (cyberbullying) is now rising as an important societal issue that affects the person emotionally and psychologically. Among the number of studies conducted to deal with cyberbullying detection, most of them are limited to only text-based content. Since the network is continuously growing, the presence of visual and audio based provoking is also needed to be considered. Therefore, the proposed study aims to develop an automatic multi-modal cyberbullying detection on social media platforms through Multi-modal Decision Fusion Classifier. The proposed multi-modal cyberbullying detection steps are data collection, multi-modality generation, score-based fusion and classification. Multiple major modalities, such as audio, visual and textual, are initially gathered from appropriate datasets. The data collected is provided as input to a multi-modal generation, where the modality is generated separately for each input. The text modality is initially generated using the hybrid Bi-directional Long Short-Term Memory assisted Attention Hierarchical Capsule Network (BiLSTM-AHCNet) model. Next, the imaging modality is generated with the Tuned Aquila EfficientB0 (Tuned AEB0) model. Finally, the audio features are extracted through the Librosa library, and the extracted features are fed to the Attention Convolutional Neural Network (ACNN) model for acoustic modality generation. In the proposed work, the Multi-Modality Decision Fusion Classifier (MMDFC) is employed to fuse all modalities, and the classification is performed by adopting the softmax layer. In addition, a weighting scheme based on horse herd optimization is applied in the fusion phase to allow accurate characterization of the features. For the simulation analysis, the proposed study used Python, and the effectiveness of the proposed techniques is analyzed by measuring several performance metrics like accuracy (98.23%), F-measure (98.22%), specificity (98.47%) and AUC (0.982).