The rise of decentralized social networks (DSNs), especially the Fediverse with platforms such as Pleroma, Mastodon, and Lemmygrad, has created new challenges for content moderation. Because each instance hosts its own data independently, there is no central authority to review posts or conversations, making it difficult to identify toxic behavior across the network.
A new arXiv preprint, titled "DeTox-Fed: Detecting Toxic Conversations in the Fediverse with Federated Graph Neural Networks," proposes a solution. The method, DeTox-Fed, applies federated graph neural networks to detect toxic conversations while keeping data distributed across instances. This allows moderation to benefit from network-wide patterns without requiring a central repository of user content.
The paper highlights the tension between the decentralized nature of the Fediverse and the need for effective moderation. By using a federated learning approach, DeTox-Fed aims to reconcile these two goals. The abstract does not provide experimental results or implementation details, but the proposed direction suggests a way to address toxic content in a privacy-preserving, decentralized manner.