Decentralized federated learning trains slower in the real world than most experiments suggest, according to new research.
Researchers mapped the math of decentralized federated learning - a privacy-preserving approach that trains models across devices without a central server - onto a well-understood physics problem: lazy random-walk diffusion on temporal networks. That mapping revealed something inconvenient. The typical experimental setup used to benchmark these systems assumes a tidy, uniform communication network. Real networks are neither. They have structural irregularities and timing gaps that most benchmarks quietly ignore.
The finding matters because decentralized federated learning is increasingly pitched as the privacy-safe alternative to centralized training - no single point of failure, no raw data leaving the device. If the convergence guarantees underpinning that pitch are built on unrealistic assumptions, the production deployments that follow will underperform those promises. The gap between lab benchmark and field result is where trust erodes.
This is not the first time federated learning research has been caught optimizing for clean experimental conditions. Earlier work on data heterogeneity across clients raised similar flags about lab-to-production gaps. The pattern here is consistent: systems that look efficient in simulation tend to hit friction when the network stops cooperating - which, in the real world, it usually does.