In today’s internet age, cyberbullying is a serious threat to people. Bullies hurt their victims by using online social media sites like Facebook, Twitter, and so on. Since then, cyberbullying or online bullying has become an increasingly significant societal problem due to its psychological and emotional impact on the victim. To overcome this issue, the proposed method is used for the identification of social network cyberbullying using many modalities. The first step is data collection, the data are collected through text, image, audio and video. Firstly, the text data is pre-processed using tokenization, stemming, lemmatization, spell correction, and PoS tagging methods. After pre-processing, feature extraction is done using Log Term Frequency-based Modified Inverse Class Frequency (LTF-MICF) and glove modelling, and the feature is extracted using the convolutional dense capsule network (Conv_DCapNet) model. Second, image data is pre-processed using Gabor filtering and image resizing. Once the pre-processing is done, the feature is extracted using image modality generation through Self-attention based bidirectional gated recurrent auto-encoder network (SA_BiGR_AENet). Third, audio (acoustic) data is extracted using the Librosa library and Attentional convolutional BiLSTM. The next step is video extraction, in this video data are extracted using keyframe extraction and Residual 101 network. Finally, in the fusion method, all four extracted features (text, image, audio and video) are fused by the deep tensor fusion framework. After that, the sigmoid layer classifies whether the data is cyberbullying or not. Then, the notification of highly detected severe is sent to the user. Thus, the proposed method detects cyberbullying. Python tool is used for implementation, and the performance metrics of recall is 96%, F1-score for bullying class 92% and weighted F1-score 91% compared with existing methods.