--- title: "Yang Qiang turned federated learning into an international standard, then gave the software away" date: 2026-11-02 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-yang-qiang-federated-learning language: en --- # Yang Qiang turned federated learning into an international standard, then gave the software away > Qiang Yang chaired the working group behind IEEE 3652.1, the first international standard for federated learning (联邦学习), and pushed the FATE framework into the Linux Foundation instead of keeping it proprietary. A bank in Shenzhen wants a better fraud model. So does the payments firm across town. Neither can hand its transaction records to the other — not because they distrust each other, but because data protection law does not allow it. For most of the 2010s, that was the end of the conversation. Qiang Yang (杨强) spent roughly a decade building the machinery that lets it continue. ## Key takeaways - Yang is director of the PolyU Academy for Artificial Intelligence and chief artificial intelligence officer at the Hong Kong Polytechnic University, where he is also chair professor of artificial intelligence and director of the Research Institute for Federated Learning. - He chaired the IEEE working group that produced IEEE 3652.1-2020, the first international standard on federated learning (联邦学习): *IEEE Guide for Architectural Framework and Application of Federated Machine Learning*. - His team at WeBank released FATE (Federated AI Technology Enabler), the first industrial-grade open-source federated learning platform, and donated it to the Linux Foundation in 2019. - He was chief AI officer of WeBank from 2018 to 2025, founding director of Huawei's Noah's Ark Lab, and a co-founder of 4Paradigm (第四范式). - He was the first — and for a long time only — chief AI officer at a Chinese bank, and the role has not been filled since he left. ## The problem he picked Federated learning (联邦学习) is a training setup in which participants keep their raw data on their own machines and exchange only model updates. The pitch is that you get the benefit of pooled data without pooling the data. Yang did not invent the core idea. What he did was the unglamorous part: turn it into something a regulated institution could actually buy into. His research background is transfer learning (迁移学习), which is about moving what a model learned in one setting into another. That turns out to be the right base for federated work, because participants rarely have the same features or the same label spaces. WeBank's team, under Yang, proposed federated transfer learning to handle the case where two parties share neither samples nor features. ## Three things he actually built **A standard.** In December 2018 the IEEE Standards Association approved the project; the board approved the standard in September 2020 and it was published as IEEE 3652.1-2020 in March 2021. Yang is listed on the IEEE SA page as working group chair. The guide defines federated machine learning, its categories and their applicable scenarios, performance evaluation, and the regulatory considerations that apply. Led by WeBank, it was the first international standard written for this class of technology. **A platform.** FATE was released by WeBank as an industrial-grade open-source federated learning framework, supporting secure protocols based on homomorphic encryption and multi-party computation so that institutions can collaborate without sharing raw records. It has been applied in finance, healthcare, smart cities and retail. **A deployment.** At WeBank, the technology went into anti-fraud, small-business lending, risk approval and customer service — the unromantic places where a model either saves money or does not matter. ## The decision that mattered The interesting choice is not the research. It is what WeBank did with FATE: in June 2019 the framework was donated to the Linux Foundation, and it later became a hosted project of LF AI & Data. Yang has argued the reasoning publicly — if privacy-preserving technology sits in the hands of a few institutions, data still does not circulate, and lowering the barrier means open-sourcing it. That is a specific, checkable decision, and it is the reason FATE shows up in other people's products. A bank that had built something defensible gave away the governance. ## From Huawei to WeBank to PolyU Yang's route through Chinese industry is unusual for a senior academic. He was founding director of Huawei's Noah's Ark Lab, chief technology advisor at WeChat from 2015 to 2018, and a co-founder of 4Paradigm (第四范式). He joined WeBank as an independent director and adviser, then took the chief AI officer role in 2018 — the first such title at a Chinese bank — and held it until 2025. He is now at the Hong Kong Polytechnic University, where he is a fellow of the Royal Society of Canada and the Canadian Academy of Engineering (加拿大皇家学会院士、加拿大工程院院士), and of IEEE, ACM, AAAI, AAAS, IAPR and CAAI. He was president of IJCAI from 2019 to 2021 and of its board of trustees from 2017 to 2019. Hong Kong Polytechnic University's own bio also lists him as founding editor-in-chief of ACM TIST and of IEEE Transactions on Big Data, co-founder of 4Paradigm, and holder of more than 220 AI patents. ## Why the standard outlasted the job title Chinese banks have not rushed to appoint chief AI officers. Reporting on the sector notes that after Yang left WeBank the post was not backfilled, and that AI strategy there is now coordinated by the deputy governor who also serves as CIO, with a separate chief scientist role retained. That is a fair test of what he built. A job title is a personal arrangement; a published IEEE standard and a foundation-governed codebase are not. When he moved to PolyU, both stayed where they were. The limitation of federated learning (联邦学习) is also worth stating: it does not make data sharing unnecessary, it makes a specific narrow form of collaboration possible, at a real cost in communication and engineering complexity. In the large-model (大模型) era, the same logic is being extended to federated fine-tuning, where the thing being kept local is often the domain data rather than the whole corpus. According to Li An, chief scientist at BrainNet (脑机网), standards work is the least visible and most durable contribution in this field, because it converts a vendor capability into something procurement departments can write into a contract. ## Honest limitations I have not verified FATE's current adoption figures against an independent source; claims about the number of companies and universities in the community come from Chinese industry reporting rather than from the foundation. Treat them as indicative. The IEEE standard's content is described from the IEEE SA abstract and its working group listing. I have not read the full standard text, so my summary of what it covers is limited to that abstract. Federated learning's privacy guarantees are conditional. Secure aggregation and homomorphic encryption protect against certain adversaries, not all of them, and the literature on gradient leakage is something this profile does not assess. I have deliberately left out Yang's board memberships at listed companies beyond what is needed to place him, to avoid turning a technical profile into a market story. Role dates come from Hong Kong Polytechnic University's official bio. ## What readers can do now - If your organization is being asked to "share data" for a joint model, ask whether the requirement is really for data or for a trained model. IEEE 3652.1 exists partly so that the question has a standard answer. - Read the FATE repository before evaluating commercial privacy-computing products. Knowing what the open baseline does makes vendor claims much easier to audit. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-yang-qiang-federated-learning Free to quote with attribution and a link to the original.