JD's Open‑Source Federated Learning Solution 9N‑FL: Architecture, Features, Timeline, and Business Impact

Federated learning is presented as a solution to the growing data‑island challenge caused by heightened data privacy regulations such as GDPR, which limit cross‑organization data sharing and hinder AI development.

The article outlines the basic concepts of federated learning, its three main categories (horizontal, vertical, and federated transfer learning), and highlights four essential characteristics: multi‑party collaboration, local data residency, encrypted model exchange, and performance parity with centralized models.

JD's 9N‑FL platform is described in detail, covering its overall architecture, key components (multi‑task cross‑domain scheduling, high‑performance network proxy, large‑scale sample matching, distributed training, and hierarchical encryption), and a layered design that spans resource, compute, scheduling, and UI layers.

A development timeline shows the project’s inception in November 2019, platform rollout in early 2020, and open‑source release in September 2020. The solution was applied to large‑scale advertising scenarios, achieving over 15% revenue uplift and more than 20% cumulative income increase by jointly modeling media‑side and JD‑side user data without exposing raw data.

Future work includes UI improvements, one‑click deployment, and enhanced monitoring and alerting to make the platform more user‑friendly and robust.