Optimizing Bullying News Detection on Social Media: An Ensemble Deep Learning Approach with Enhanced Optimization Algorithm

Cyberbullying is a growing digital threat that exploits online platforms to harm individuals, which can take place on social media, messaging environments, gaming, and mobile phones. Cyberbullying can result in deep psychiatric and emotional illnesses for those affected. Henceforth, there is a crucial necessity to develop an automated model for cyberbullying detection. Detecting cyberbullying is very challenging, due to the occurrence of complete affective content, which is also frequently sarcastic, and multimodal, such as audio, text, and image. Presently, Deep Learning (DL) techniques have attained extraordinary achievements in numerous tasks and are employed for the detection of cyberbullying for multimodal data. This paper presents the Multi-Layer Perceptron and Feature Extraction for Detecting Bullying in Multimodal Content (MLPFE-DBMMC) technique. The MLPFE-DBMMC technique aims to develop an effective multimodal cyberbullying detection framework by leveraging audio, textual, and visual inputs to recognize and identify harmful behaviour. Initially, the preprocessing stage is performed in multimodal formats such as image, audio, and text. For image preprocessing, the anisotropic diffusion method is employed for noise removal. The stationary wavelet transform-based noise removal is applied in the audio preprocessing. Moreover, the text preprocessing stage involves various levels such as lower case, tokenization, removal of stopwords, and stemming. For the feature extraction process, the MLPFE-DBMMC model implements the Contrastive Language-Image Pre-Training (CLIP) method for image, VGGish-based audio, and the generative pre-training-2 (GPT-2) technique is employed for text. After feature extraction from the multimodal data, output features of size (3672, 512) for images, (3672, 128) for audio, and (3672, 768) for text are obtained. Three multimodal features are fused via concatenation for final classification. The MLPFE-DBMMC model implements the correlation alignment loss (CORAL) multi-layer perceptron technique for the multimodal cyberbullying detection process. Finally, the improved artificial rabbit’s optimization (IARO)-based hyperparameter tuning is performed to enhance the detection outcomes. A wide range of experiments of the MLPFE-DBMMC approach is performed under the MultiBully dataset. The comparison study of the MLPFE-DBMMC approach portrayed a superior accuracy value of 94.44% over existing techniques.