Researchers Propose Byzantine-Robust Federated Learning for Fire Detection

A research team has published a preprint exploring the use of federated learning (FL) for indoor fire detection systems. These setups rely on edge cameras recording sensitive video feeds that cannot easily be aggregated onto a single central server.

To overcome communication bottlenecks and security risks, the researchers introduced a curated dataset from eight public sources and an edge detector that compresses model updates up to 10 times. They also developed a semi-decentralized Byzantine-robust FL approach combining history-aware aggregation with a rotating coordinator, effectively eliminating single points of failure and filtering out stealthy attacks.

Evaluation on a held-out test set and a distributed six-node cloud deployment confirmed that the proposed framework matches fixed-server configurations in both detection speed and balanced accuracy.