JPMorgan Kinexys and BNY jointly developed Project AIKYA, a federated learning (FL) proof-of-concept for cross-institution anomaly detection in financial transactions, published on August 11, 2025. Note: this is a PoC using synthetically generated data inside a permissioned, controlled network — not a production deployment, and “possible benefits are illustrative; outcomes are not guaranteed” per the source.
Although machine learning models have evolved significantly, their training data is typically siloed within individual banks. Cross-border and cross-institution anomaly detection — where insights are distributed across geographies and institutions — remains constrained by legal, regulatory, and competitive considerations. Centralized data repositories spanning multiple institutions are largely impractical. Federated learning emerged as a class of techniques that move away from a single central entity toward distributed systems and privacy-enhancing techniques.
Project AIKYA was published August 11, 2025 by Kinexys by J.P. Morgan, documenting collaboration with BNY in a permissioned, controlled network with trusted participants using synthetically generated data. Following the PoC, the source identifies several directions to explore: real-world data validation, scaling to multi-participant networks, FL models that support participants with heterogeneous technology stacks and/or model-agnostic networks, and testing across multi-region and multi-industry networks.
The PoC evaluated FL as a privacy-preserving mechanism for institutional cooperation. Each participant trained a local model on private data without exposing raw information; only model weights were exchanged and aggregated. Pre-trained models demonstrated specialized anomaly detection capabilities — one bank's model excelled at location-based anomalies, the other at account-age-based anomalies, while each was poor on the other's specialty, demonstrating a lack of generalization to novel patterns. A single round of aggregation caused a significant decline in predictive capability, attributed to aggregation shifting model weights without sufficiently integrating broader anomaly knowledge. After round-2 aggregation and retraining, performance began recovering; by round 4, the aggregated model converged, achieving robust detection of both anomaly types across both client datasets.
PoC scale only. Permissioned, controlled network with trusted participants. Synthetically generated data used to simulate real-world scenarios. The federated model achieved comparable or superior anomaly-detection coverage to local models without exposing any raw data, validating FL as a viable approach for cross-border payments and complex financial transactions.
Details