JD Digital Science Unveils Fast Secure Federated Learning Framework and Two Industry‑First Techniques

Recently, JD Digital Science Group AI Lab chief scientist Bo Liefeng disclosed JD's federated learning strategy and introduced its fast secure federated learning framework, sharing two industry‑first breakthroughs: a kernel‑based nonlinear federated learning algorithm and a distributed fast homomorphic encryption technique, both accepted at KDD 2020.

At an AI finance open class hosted by the Hong Kong AI & Robotics Society, Bo noted that data‑privacy regulations such as the EU GDPR and China’s draft Data Security Management Measures pose challenges for AI deployment, and federated learning emerges to address privacy and security from the source.

Federated learning enables AI systems to jointly use data while satisfying privacy, security, and regulatory requirements, allowing multiple enterprises to train models without sharing raw data.

JD Digital Science has applied federated learning to face‑recognition scenarios, achieving a 99.96% pass rate at a 1e‑5 false‑positive rate and 99.99% at a 1e‑4 rate, demonstrating superior model performance through multi‑party data collaboration.

Industry challenges include gradient‑based attacks, efficiency loss due to gradient encryption, and resource waste from synchronous updates; JD’s fast secure framework addresses these with three features: privacy‑preserving intermediate value exchange with perturbation, centralized data exchange independent of participants, and an asynchronous computation model that greatly speeds training.

The framework incorporates the kernel‑based nonlinear federated learning algorithm using doubly stochastic gradient descent, which avoids transmitting raw samples or gradients and improves speed, and a distributed fast homomorphic encryption technique for large‑scale secure computation; both resulted in papers accepted at KDD 2020.

AI is a key pillar of new infrastructure; JD Digital Science’s Industrial AI Center consolidates multiple AI labs to industrialize frontier AI technologies such as federated learning and to provide AI capabilities as infrastructure for digital transformation.