Empowering privacy and resilience: a decentralized federated learning approach to cyberbullying detection

In a rapidly changing digital world, the rise of cyberbullying has become a pressing issue that calls for creative and flexible solutions to detect and prevent it. To address this urgent need, we introduce a novel method: decentralized federated learning for cyberbullying detection using a ring topology network. Traditional federated learning (FL) paradigms have traditionally relied on centralized servers to manage important operations. However, this centralized model faces complex challenges related to privacy vulnerabilities, fairness disparities, and scalability constraints. Our solution brings about a significant change, embracing decentralization as the foundation of a strong and privacy-focused cyberbullying detection framework. This research represents a significant shift away from centralization, as it relies on distributed clients in a ring topology network to collectively carry out crucial FL tasks. Through the redistribution of these responsibilities and the establishment of direct communication channels between neighboring nodes, our approach successfully avoids the challenges related to central server instability and bias. The ring topology architecture creates a decentralized ecosystem, carefully crafted to prioritize the confidentiality of user data while enhancing the resilience and efficiency of the FL process. Our model architecture for cyberbullying detection is carefully crafted, utilizing the powerful capabilities of GRU, LSTM, BERT, and Word2Vec embedding, along with emotional features. This architectural innovation perfectly aligns with the ring topology FL approach, enabling localized updates, efficient aggregation, and flexible adaptability. It is worth mentioning that the BERT model consistently outperforms its competitors, delivering exceptional results.