The internet’s evolution as a global communication nexus has enabled unprecedented connectivity, allowing users to share information, media, and personal updates across social platforms. However, these platforms also amplify risks such as cyberbullying, cyberstalking, and other forms of online abuse. Cyberbullying, in particular, causes significant psychological harm, disproportionately affecting young users and females. This work leverages recent advances in Natural Language Processing (NLP) to design a robust and privacy-preserving framework for detecting abusive language on social media. The proposed approach integrates ensemble federated learning (EFL) and transfer learning (TL), combined with differential privacy (DP), to safeguard user data by enabling decentralized training without direct exposure of raw content. To enhance transparency, Explainable AI (XAI) methods, such as Local Interpretable Model-agnostic Explanations (LIME), are employed to clarify model decisions and build stakeholder trust. Experiments on a balanced benchmark dataset demonstrate strong performance, achieving 98.19% baseline accuracy and 96.37% with FL and DP respectively. While these results confirm the promise of the framework, we acknowledge that performance may differ under naturally imbalanced, noisy, and large-scale real-world settings. Overall, this study introduces a comprehensive framework that balances accuracy, privacy, and interpretability, offering a step toward safer and more accountable social networks.
Kumari et al. (2025) studied this question.