Federal learning has applications in industries such as finance, healthcare, smart manufacturing, smart home, and smart travel.

In recent years, data privacy protection has received increasing attention. In June 2017, China began to implement the “Network Security Law of the People’s Republic of China”, which stated that “ network operators must not disclose, tamper with, or damage the personal information they collect; without the consent of the collector, Provide personal information to others. In 2018, the European Union introduced the first General Data Protection Regulation (GDPR) on data privacy protection, which clarified certain provisions on data privacy protection.

This means that the collection of user data must be open and transparent, and data cannot be exchanged between enterprises and organizations without user authorization.

The other side is the problem of “data islands” faced by artificial intelligence applications. Each company or organization holds a limited amount of data and has its own characteristics. Giant companies monopolize large amounts of data, and small companies have difficulty obtaining data. Due to factors such as competition, security issues, and approval processes, data is in different owners. There are hard-to-break barriers in the circulation between the parties, forming a “data island”.

In the face of “data compliance” and “data silos”, Google Inc. first proposed the “Federated Learning” algorithm framework. To put it simply, this technology enables participants to build models without revealing the underlying data. can be safely compliant under Solve the problem of information silos and complete common modeling.

In China, the AI ​​team of Weizhong Bank proposed a system-based solution based on “Federal Learning”. At the 23rd CCF TF seminar on October 26th, Fan Lixin, chief scientist of Weizhong Bank’s artificial intelligence, told: “Federal learning conditions that do not leak raw data Under the joint modeling, the rules, knowledge and value behind the data are fully utilized.For Micro Bank, Federal Learning is a open source ecological platform, not a single technology. On the platform we will be different application scenarios provides a variety of support.

At the meeting of the day, fromWeizhong Bank, Tencent Cloud, Huawei, VMware China R&D Center, Jingdong Smart City Division, Innovation Workshop, Ping An Technology, the ubiquitous computing system research center of the Institute of Computing Technology of the Chinese Academy of Sciences, and the Guanghua School of Management of Peking University shared the latest application case of AI federal learning.

Financial Area

In the financial sector, the use of federated learning joint modeling wind control model can more accurately identify credit risk and joint anti-fraud; and the federal anti-money laundering model established by federal learning can solve the problem of small sample and low data quality in this field. problem.

To do a good job of credit risk control for SMEs, you need a central bank’s credit report, as well as upstream and downstream tax, business and other data. Most small and micro enterprises only have the central bank’s credit report, tax and financial data are not available, and vertical federal learning can provide a good solution. The intelligent scoring engine developed by Weizhong Bank AI team can jointly model the invoiced amount and the central bank’s credit data and other label attributes to reduce the loan non-performing rate of small and micro enterprises by 2%.

AML is playing an important role in the daily operations of the bank. Effective anti-money laundering activities can curb economic criminal activities. In the past, the bank’s approach was to use the traditional rule model to filter the apparent non-money laundering records and manually view the remaining records. The scope of the model coverage is small, and it still takes a lot of time in manual review. The federal anti-money laundering model established by federal learning can solve the problem of small sample and low data quality in this field, and shows the monitoring efficiency and effectiveness level that the traditional rule model cannot achieve.

Medical Area

In the medical field, everyone’s medical data is absolutely private. And In fact, there are many diseases that require data modeling. As a result, federal learning has become a good solution. At the seminar, Prof. Chen Yiqiang, director of the Research Center for Ubiquitous Computing Systems of the Institute of Computing Technology of the Chinese Academy of Sciences, cited the case of medical integration for Parkinson’s disease.

“One of the hallmarks of Parkinson’s disease is that medication intervention is very effective for Parkinson’s patients.But the variety, time and amount of medication must be strictly prescribed. How to establish a model that can turn the original doctor’s qualitative drug delivery into a quantitative and objective method, how to make the patient use this model at home, this is a big challenge.” Chen Yiqiang said, < /span>“So we built a FedHThe ealth framework uses a federated learning to migration process to learn incremental learning, stringing the entire model.

On the one hand, in the hospital scene, the effective cycle of the efficacy of the drug is recorded by the doctor; on the other hand, the life of the Parkinson patient is measured by the wearable device in life, and finally the model is established to combine the data of the two parties. , measuring MDS-UPDRS, this internationally popular Parkinson scale, to achieve medical integration.

Through this model, it is possible to judge the condition of the patient before and after taking the drug, to determine whether the drug is good or not, the dose of the drug is wrong, which drugs have effects on the upper limbs, and which drugs may have poor effects on the lower limbs.

Smart City Construction

Resolving urban pain points by continuously acquiring, integrating and mining big data in different areas of the city is the way for today’s cities to lead to smart cities. Zhang Yibo, head of the AI ​​platform department of Jingdong Smart City Division and senior researcher of Jingdong Smart City Research Institute, shared the construction of a credit city system based on big data and federal learning, and the product of Jingdong City based on urban computing and federal learning technology – digital gateway .

To solve the problem of data islands and data sharing difficulties in the city, create secure, shared, intelligent and efficient connections between different government agencies, large enterprises and institutions, Internet companies, etc. Digital gateways use federal learning technology. Based on its advantages of security, credibility, non-destructiveness, diverse scenarios, ease of use, lightweight deployment, and trusted distribution, it helps inter-domain modeling and use under the premise of legal compliance.

Smart Terminal

The form of smart terminals is increasingly moving towards distributed AI. Hua Wei, CEO of CTO Office of Huawei Consumer BG Software Department, said at the seminar that the core value of distributed AI lies in accurate perception and accurate prediction, and it faces system dynamics, heterogeneous devices, multi-end multi-user collaboration, and adaptation. The challenge of hardware features. Among them, breaking through the system dynamics, equipment heterogeneous barriers, building a unified feature space, and based on the unified feature space for multi-user and multi-device collaborative training, to bring users a unified, continuous, personalized service experience is A potential opportunity for federal learning.

Cloud Service

Zhang Haining, Technical Director of VMware China R&D Center, believes that cloud service is an ideal way to learn from the federal government. The federal learning has its own characteristics, suitable for deployment and use on the cloud and multiple users, for example, The organization of federal learning in the public cloud is added to form a heterogeneous system or ecosystem, providing a platform for data docking between different organizations.

Federal learning is used in smart manufacturing, smart home, smart travel and other industries, and the landing case is gradually enriched . For which application is availableThe problem of being able to run faster, Fan Lixin, chief scientist of Weizhong Bank’s artificial intelligence, said in an interview at the meeting, “We don’t make this prejudgment. The reason why we do this federal learning ecological platform is for everyone. Being able to participate is to test the market and practice. Once a good sign comes out, smart people and capital will invest in these directions, and then continue to make breakthroughs.The owner of the data needs this data. The integration of data, the needs of data, and the participation of regulators. At present, the technical details of the practical application of federal learning are still perfect and rich, and it is necessary to have a ‘first mover’ to enter the market and cultivate the market.”