Abstract
<title>Abstract</title> <p>With the fast expansion of social media sites and multilingual social media conversations , the problem of cyberbullying is becoming a huge socio-psychological problem across the globe. In general, the cyberbullying detection techniques are mostly dependent on data gathering and processing in centralized form, which makes the system vulnerable to privacy breaches and data leaks. This paper proposes a framework for multilingual cyberbullying detection that ensures privacy and security by combining NLP techniques, deep learning models, federated learning (FL), and blockchain technology. Our proposed framework will be processing the social media text in English, Bangla, and Banglish by means of text normalization, emojis extraction, tokenization, and Personal Identifiable Information (PII) masking. With the help of federated learning, the collaborative training of model will be possible without any exchange of user data, which guarantees the privacy of users. In order to ensure security and transparency in the process, blockchain technologies such as SHA-256 hashing, Hash-based Message Authentication Code (HMAC) validation, Merkle Root check, and Proof-of-Work (PoW) consensus will be used in the process. Four classification models such as Bidirectional Long Short-Term Memory (BiLSTM), Long Short-Term Memory (LSTM), Random Forest, and Extreme Gradient Boosting (XGBoost) have been implemented and tested for checking the efficiency of our proposed framework. As a result, it has been found that the model built with the BiLSTM technique 1 showed the highest level of accuracy of 98.43%, whereas other three models provided the accuracy of 94.43% (LSTM), 93.20% (XGBoost), and 81.76% (Random Forest).</p>