DeTox-Fed uses federated GNNs to detect toxic chats across Mastodon instances
New federated learning framework spots toxic conversations while keeping user data private on decentralized social networks.
Deep Dive
DeTox-Fed is a federated graph-learning framework for detecting toxic conversations in the Fediverse (Pleroma, Mastodon). Each instance builds a local conversation graph and trains a Graph Neural Network collaboratively without sharing raw conversations or labels. Tested on a large Pleroma dataset, it achieves stable detection under limited labels and partial client participation, enabling semi-automated moderation in decentralized social networks.
Key Points
- DeTox-Fed uses federated Graph Neural Networks (GNNs) trained across decentralized instances without sharing raw conversations or moderation labels.
- Each instance builds local conversation graphs where nodes represent threads and edges capture shared user participation—enabling structural toxicity detection.
- Evaluated on a large Pleroma dataset, the framework maintains stable detection even with limited labels, partial client participation, and varying moderation thresholds.
Why It Matters
Enables privacy-preserving toxicity moderation at scale for the Fediverse, balancing autonomy and safety without centralizing user data.