Federated Layer-Wise Exchange for Cross-Client Money Laundering Pattern Detection
The framework introduces a per-step, layer-wise embedding exchange protocol to recover centralized node representations in federated Graph Neural Networks. This method synchronizes intermediate node representations across clients at every layer of the forward pass, effectively closing the 'representation-equivalence gap' caused by missing remote node information. Empirical results show that combining this exchange with parameter aggregation is superior to using exchange alone.