Baidu's Interoperability Solutions for Federated Learning: Principles, JinKe Alliance, and the Open‑Source HIGHFLIP Protocol

This article shares Baidu's thinking and practice on federated‑learning interoperability.

Main contents include:

Introduction to interoperability principles and schemes.

Status and progress of the JinKe Technology Alliance solution.

The open‑source collaborative federated HIGHFLIP scheme.

Discussion and reflections on the two types of solutions.

1. Interoperability Principles and Schemes Baidu analyzed various federated‑learning platforms, classifying them into four layers (application, scheduling, algorithm, security operator). Based on this analysis, Baidu proposes three interoperability models—white‑box, gray‑box, and black‑box—and further classifies them into bottom‑layer and top‑layer approaches.

Bottom‑layer interoperability aligns each layer’s protocols (application, scheduling, algorithm, operator) between heterogeneous systems, while top‑layer interoperability abstracts the entire platform and only aligns a high‑level interface, reducing integration complexity.

2. JinKe Technology Alliance Solution Baidu contributes the scheduling layer and algorithm‑container loading for the JinKe Alliance. The solution defines clear communication for each layer, adopts transient containers as the primary model, and minimizes coupling by using self‑describing metadata, labels, and environment variables.

3. Open‑Source Collaborative Federated HIGHFLIP Scheme HIGHFLIP is a top‑layer communication protocol that leverages gRPC and ProtoBuf for standardized messaging and ONNX for model representation. It requires only an adapter and a plugin to interconnect heterogeneous federated‑learning platforms, offering a low‑cost, weakly invasive integration path suitable for commercial products.

Key features of HIGHFLIP include:

Standardized gRPC/ProtoBuf communication.

Model compatibility via ONNX.

Bridge architecture that translates between heterogeneous nodes without deep system changes.

Support for combining operators from different platforms in a single workflow.

The protocol is being co‑maintained by Baidu, FATE, Tencent, and JD, with open‑source release planned on GitHub in 2023 and integration into FATE 1.9.

4. Discussion and Reflections The bottom‑layer solution offers fine‑grained control but requires extensive cooperation among vendors, while the top‑layer HIGHFLIP protocol provides rapid deployment with minimal intrusion. Both approaches are complementary and can be combined to achieve a complete interoperability ecosystem.

In summary, Baidu's work spans both deep, layer‑by‑layer integration and lightweight, top‑level protocol design, aiming to facilitate seamless collaboration across diverse federated‑learning platforms.