No abstract

Cyberbullying has become one of the most pressing concerns for online platforms, putting individuals at risk and raising severe public concerns. Recent studies have shown a significant correlation between declining mental health and cyberbullying. Automated detection offers a great solution to this problem; however, the sensitivity of client-data becomes a concern during data collection, and as such, access may be restricted. This paper demonstrates FedBully, a federated approach for cyberbullying detection using sentence encoders for feature extraction. This paper introduces concepts of secure aggregation to ensure client privacy in a cross-device learning system. Optimal hyper-parameters were studied through comprehensive experiments, and a computationally and communicationally inexpensive network is proposed. Experiments reveal promising results with up to 93% classification AUC (Area Under the Curve) using only dense networks to fine-tune sentence embeddings on IID datasets and 91% AUC on non-IID datasets, where IID refers to Independent and Identically Distributed data. The analysis also shows that data independence profoundly impacts network performance, with AUC decreasing by a mean of 5.1% between Non-IID and IID. A rich and extensive study has also been performed on client network size and secure aggregation protocols, which prove the robustness and practicality of the proposed model. The novel approach presented offers an efficient and practical solution to training a cross-device cyberbullying detector while ensuring client-privacy.

No abstract

No abstract

No abstract