As a new technology, “federated learning” can jointly build the model and improve the effectiveness of the AI model based on guaranteeing data privacy security. Under this prospect, since “federated learning” has been carried out, it gets attention from parties. In China, Professor Qiang Yang, WeBank chief artificial intelligence officer, President of The Association for the Advancement of Artificial Intelligence and WeBank’s AI team first released the FATE (Federated AI Technology Enabler), a commercial-level open-source project, for the popularization and implementation of “federal learning”.
As the first open-source community of “federated learning”, FATE attracts the attention of a large number of technology enthusiasts and university research teams, including Qinghe Jing, a graduate student from The Hong Kong University of Science and Technology (HKUST). As the first contributor, he proposed optimization suggestions for the FATE communication framework which significantly increased communication efficiency.
For this, we spoke to Qinghe Jing about his and his team’s research on FATE, as well as their expectations and assumptions about FATE.
1.The first well implementation for “federated learning” in China
One of the main tasks of the HKUST research team, where Qinghe Jing is based, was to optimize the machine learning framework. Through this opportunity, Qinghe Jing learns the new concert about “federated learning”.
Essentially, “Federated Learning” is a kind of distributed machine learning technology based on multiparty secure computation. It allows participants to jointly build models without disclosing the underlying data and the encryption (obfuscation) patterns of the underlying data. As a way of win-win machine learning, it can effectively connect “data silos” to form an AI continent. In industrial applications, it helps different institutions break down barriers and jointly build AI models. Users can achieve the best privacy protection while the parties’ data can be never left local storage.
However, FATE as the implementation project of “federated learning”, it was a focus by the research team. After the deep study and discussion of FATE, a group of young technical enthusiasts begins to try to use FATE to solve the problem in research. Therefore, they provide optimization suggestions for the FATE open-source project.
According to Qinghe Jing, after contacting with FATE for days, they think this is the first “well implementation” of “federated learning” in China. In their point of view, FATE enriches and expands the concept of “federated learning”. FATE not only includes horizontal federated learning, but it also contains vertical federated learning. Moreover, it combines the “federated learning” and “transfer learning” so that FATE can enable different organizations and patterns of data to collaborate. In their opinion, such an implementation form enables FATE to be chance to provide cooperation for those institutions which have a special requirement like two companies that have concerns about users’ privacy protection.
2.Based on “federated transfer learning”, suitable scenarios of FATE are wide
According to Qinghe Jing’s thought, FATE can be applied to a wide range of scenarios, especially in finance, medical care, etc. In a more data-sensitive and regulated scenario, FATE can help achieve collaboration while ensuring data privacy protection.
Qinghe Jing and his research team understand that it’s impossible to have the same structure information when two units train models during the real usage scenario, which nearly can’t happen. The application scenarios would be quite limited even if it were possible to restrict the entire user to have the same model. Therefore, combined with vertical federated learning and federated transfer learning, FATE can further expand the applicable scenarios. Even the models of two organizations are different, FATE still can train together using the shared data. This will be also appropriated by young researchers.
For example, WeBank and several Banks jointly build the anti-money laundering model under the premise of not sharing users’ data. Through the model testing, the more participants of the bank, the performance of the model is higher. As for the medical care scenario, such as medical imaging and medical cases are scattered in multi hospitals, federated learning can improve the accuracy of disease prediction and the overall medical diagnosis and treatment level.
Except for the ability of study and model building, when we are facing the coming era of 5G, internet security, data security is especially the main issue. Everyone is talking about putting a safety valve on data. From the General Data Protection Regulation (GDPR) issued by the European Union to the draft of Data security management measures (consultation draft) drafted by The Cyberspace Administration of China, the free transmission of data under the premise of safety compliance has become the general trend.
In Qinghe Jing’s point of view, FATE can optimize encryption algorithms by using the homomorphic encryption, not DP and other forms so as to strongly protect the data security and have a significant promotion on the AI implementation.
3.FATE in the minds of young research teams and technology enthusiasts
As the first contributor of the FATE open-source project, he is also a technology enthusiast who knows and uses FATE early. Qinghe Jing and his research team’s fellows have a lot of images of construction for FATE. They believe that FATE framework language can achieve more model applying in different scenarios. As for the performances, they will continue to research deeply to explore the possibility of optimization.
At the end of the interview, Qinghe Jing said, FATE is a very good and new thing, providing a bigger help for AI implementation. It also enables AI to apply in more suitable scenarios and make the model studied by machine learning more reliable in different fields, which can let machine learning enter our life better. In the future, he will keep following FATE and expect more technology enthusiasts and he can finish more significant jobs by the updated framework.
4.The open-source ecosystem work with ambition technology enthusiasts
It is reported that Qinghe Jing is from HKUST SingLab, Kai Chen, the associate professor of HKUST, director of SingLab, leads students to jointly research the design, analysis and implementation of a network system. The research can address the communication challenges of AI facility Scale Out to improve the computing power of AI basic infrastructure and enable the federated learning technology to implement better and faster.
Contributors like Qinghe Jing has a good beginning. We can expect that more and more universities and young technology enthusiasts will join the construction of the FATE open-source project. Professor Qiang Yang said that a dynamic open-source ecosystem cannot live without the contributions of aspiring young people.
Professor Qiang Yang points out that the safety and privacy protection of intelligent computing is one of the most important development direction of the software industry. FATE has become the main force of sustaining big data, artificial intelligence and multi-party computing architectures. We are welcome more technology people like Qinghe to participate in the construction, making the ecosystem of “the operation system of big data learned by machine” more and more dynamic.
In the meantime, when facing the federated learning development concerned by a lot of technology enthusiasts and developers and the future development of this technology, professor Qiang Yang said, “federated learning is an effective technology of protecting users’ data privacy. Its research and industry application cannot be separated. At the next development, we expect more and more companies and institutions can adopt federated learning technology and develop the AI vertical application of 2C and 2B
The society requirement of privacy and data security is stricter. It will promote the development of new encryption technology and multiple model building. Let’s expect it!