Kinexys, BNY, RBC Launch Federated-Learning Fraud Detection with DeepTempo and NVIDIA

J.P. Morgan‘s Kinexys blockchain unit has brought BNY, Royal Bank of Canada, DeepTempo, and NVIDIA into a collaborative artificial-intelligence project for fraud detection. The initiative applies federated learning, allowing institutions to train shared models on proprietary transaction data without pooling the underlying records.

The architecture addresses a structural tension in cross-bank fraud work: individual datasets are too small or too skewed to catch novel patterns, yet regulatory and competitive constraints prevent centralised pooling. Federated learning keeps data on-premise while aggregating model weights, a design that scales detection power without creating a new honeypot or breaching data-residency rules.

No launch date, transaction volume, or detection-accuracy benchmark appears in the announcement. The initiative stands as a proof-of-concept until one of the participating banks confirms production traffic or a regulator references the framework in guidance on model risk or cross-border information sharing.