Impact of network topology on Byzantine resilience in decentralized federated learning

Federated learning trains a model across many machines without moving their data. Take away the central server and you have decentralized federated learning: peers exchange updates directly, over whatever network happens to connect them. The question this paper asks is simple. When some of those peers lie, does the shape of the network decide whether the honest ones can still learn?

The gap

Byzantine-robust aggregation rules exist to tolerate nodes that deviate from the protocol, whether by fault or by attack. They have been evaluated almost exclusively in two settings: a client-server star, or a fully connected graph where every node hears every other. Neither is what a real peer-to-peer deployment looks like. The literature had not yet tested whether robustness guarantees survive contact with sparse, large, irregular topologies.

What was done

The team built a framework for running decentralized training over arbitrary graph structures and used it to place two state-of-the-art robust aggregators, Krum and GeoMed, inside small-world and scale-free networks with Byzantine nodes present. Byzantine nodes were placed two ways: at random, and strategically, on the rewired long-range edges of a small-world graph or on the highest-degree hubs of a scale-free one. The instrument grew out of FedUp, a peer-to-peer training platform first built at SpartaHack 2024.

Finding

Random placement barely distinguishes the topologies: honest accuracy falls from 95% to 87% at 30% Byzantine nodes and collapses at 60%, on both graphs and with either aggregator. Strategic placement does. A small-world network still exceeds 90% accuracy with 27% of nodes Byzantine, because its edge density is uniform and no small set of nodes controls connectivity. A scale-free network fails to converge once a quarter of its hubs are poisoned, at the same proportion it tolerates at random. Robustness, as currently defined and measured, is a property of the aggregator and the graph together, not of the aggregator alone.

Why it matters

Any real deployment of decentralized learning will run over a sparse network, and real networks have hubs. The paper’s conclusion is a direction for the field rather than a fix: aggregation schemes need to be topology-aware before they can be trusted at scale.