Shulou(Shulou.com)06/02 Report--

2019-09-29 11:43:40

Author | Just

Produced | AI Technology Camp (ID:rgznai100)

There is no doubt that in the period when the industry is looking forward to the landing of artificial intelligence (AI) applications, the important fulcrum of data has increasingly become a "choking" problem.

AI needs data to optimize the effect of the model, but most enterprises will not easily provide data to AI unconditionally, because data is their survival card to some extent, which also leads to a situation in which a small number of giant companies monopolize a large amount of data, while it is difficult for small companies to obtain data. On the other hand, data fusion is even more difficult because of the provisions of laws and regulations on data privacy protection. The problem of data isolated island seems to be an inextricable knot, and the landing process of artificial intelligence has been seriously hindered.

At this time, the Federal Learning (Federated Learning) technology proposed by Google in 2016 began to be highly expected in the industry. Domestic companies represented by WeBank, Ping an Technology and Baidu have become "early adopters" of new technology, hoping that it will become a bridge to open up the isolated island of data.

Google took the lead in establishing a federated learning system to solve the data privacy problem of users' personal terminal devices. In Android mobile phone users, the initialization model is first downloaded to each terminal, and then the model parameters are updated according to their own data. Different terminals then produce different update results and send them to the cloud for aggregation. The summarized model parameters will be used as the initial parameters of the next update and will be iterated until convergence.

In this way, we can not only ensure that data are not shared and protect users' privacy, but also share a general model that can be constantly updated in the cloud using swarm intelligence. This is federated learning technology, and the related technologies with this technology as the core are collectively referred to as federated intelligence. Dr. Wang Jianzong, deputy chief engineer of Ping an Technology and head of the federal learning team, put forward this concept for the first time. In a recent interview with AI technology stronghold (ID:rgznai100) and other media, he said that federal learning is to federal intelligence as deep learning is to artificial intelligence, but federal intelligence is still in the category of artificial intelligence, and its ultimate goal is to achieve artificial intelligence.

Breaking the data isolated Island, the Application and practice of Federal Learning

As an early user of federated learning technology, Google opened up the federated learning framework TensorFlow Federated in February this year, which can be used for machine learning and computing experiments of decentralized data. In China, WeBank AI team opened up the self-developed "Federal Learning FATE (Federated AI Technology Enabler)" learning framework, and promoted its application in credit risk control, regulatory technology and other fields.

Ping an Technology is also independent of the Honeycomb federal learning platform, and has a relatively mature landing case. Wang Jianzong mentioned that based on the financial and insurance business data of Ping an Technology, they jointly model the data that could not be obtained before through federal learning technology. as a result, it can accurately predict the overdue default rate of user loans or credit cards, predict the purchase behavior of cross-domain products, and predict insurance customers through bank customers, except in the fields of finance and insurance. They have also carried out practical applications in many fields, such as medical treatment, intelligent voice and vehicle networking.

What these applications have in common is that federal intelligent solutions require real-time encryption and decryption of data during transmission, and efficient implementation on this basis, such as deep learning and training. to achieve tens of millions or even hundreds of millions of parameters exchange, synchronous, asynchronous processing. Based on this, the new scheme should not only ensure the security of multi-source data in the tuning process of AI model, but also effectively evaluate the contribution of each data source to the final optimization result.

In short, in order to achieve joint modeling in a distributed environment, it will naturally put forward corresponding requirements for hardware support, and the cooperation between Ping an Technology and Intel provides an effective solution to the above problems.

At the hardware level, the two sides described cooperation at the federal learning technology level as a "hit-and-hit". Intel has always wanted to do a trusted computing data analysis execution environment, hoping to effectively prevent external access and attacks on sensitive data and applications. Intel's newly released SGX (Software Protection extension) technology achieves this, which uses processor instructions to create trusted areas in different data sources for data access, which meets the current requirements of federated learning computing.

Wang Jianzong said that the initial configuration of SGX was not created for federation learning, but the hardware trusted platform gradually opened some special interfaces in the later stage, so that the interfaces can be directly encapsulated, thus faster and more efficient in the process of encryption and decryption of information transmission. This method of "hardening" the trusted computing environment can speed up iterative training, and is also in line with the current trend of software hardening and hardware softening.

Compared with the traditional soft encryption methods, such as reforming the traditional deep learning framework TensorFlow, PyTortch, Caffe and MxNet, the encryption and decryption process in information processing and transmission will consume too much time.

Intel ®SGX Technology enhances data Security with trusted "enclave"

Specifically, Intel ®SGX technology can help prevent internal and external attacks by constructing a credible "Enclave" in specific hardware (such as memory) for the interaction and transmission of intermediate parameters, so that the security boundaries of data and applications are limited to the enclave itself and the processor, while its operation does not depend on other hardware and software devices. This means that the security protection of data is independent of the software operating system or hardware configuration, even if the hardware drivers, virtual machines and even the operating system are attacked and destroyed, it can more effectively prevent data leakage.

Federal Learning Program for Intel ®SGX Technology

Based on the characteristics of Intel ®SGX technology, the federated learning team, together with Intel, designed a 1N multi-source data AI model training method in its federated learning scheme, which is helpful to accurately evaluate the contribution of each node data to AI model training and facilitate users to adjust the scheme.

Take the application of federal learning in the insurance industry as an example. In the past, when users applied for insurance, business staff could only determine the premium amount according to the user's age, gender and other basic information. However, with the continuous development of the information society, the number and characteristic dimensions of user data have been greatly increased. For example, for health insurance, if the business system can use massive medical records, family medical history data and other data to predict AI. And get a more subdivided category of health assessment, which is expected to improve the accuracy of the health assessment results of policyholders.

Among them, medical records and medical history are undoubtedly the data that absolutely need to ensure privacy in various health care institutions, which are not only impossible to be made public, but also need to be protected by upgrading the level of security. Now with the introduction of federal learning program, insurance companies can carry out AI training of insurance pricing model without touching user data. From the current effect, federal learning solution has significantly improved the effect of personalized insurance pricing.

Of course, the application of new technologies is always accompanied by new challenges, and federal learning has its own shortcomings that cannot be solved at present. Wang Jianzong pointed out that at present, federation learning uses different algorithms to transform models for joint modeling, and there is no tool or methodology that can solve the problem of federalization of all deep learning algorithms.

At the same time, different from the decentralized mechanism of blockchain, federal learning forms a centralized federal government. There is only one common model to distribute in the "federal government", so the problem of "two trusts" needs to be solved: one is to ensure that there is a federal government that all participants trust, and the other is that the operation information of the federal government should be transparent.

However, various new technologies are constantly evolving, and Wang Jianzong believes that as long as more enterprises and practitioners join the team of using federal learning, these problems will be solved gradually.

He compared the dilemma he faced when working on information systems more than a decade ago. At that time, the development language of each information system was different, but now it has been completely solved; the problem of data isolated island still exists, but then the emergence of cloud computing makes the situation that hundreds of systems do not communicate with each other, so it also makes him optimistic about the development prospects of federal learning in the future.

Federal Intelligence, leading the New Dawn of AI Innovation?

The bigger game of federal learning technology is for participants to work together to create a federal learning ecology, but Wang Jianzong said that the most important thing now is to seize the opportunity and come up with ecological and solutions for the future trend of federal learning.

Ecology is inseparable from the layout of the system architecture. At the hardware level, the current cooperation between Intel and Ping an Technology is still in the first stage, only taking out a hardware encryption box to solve one of the links of data training. The Ping an Science and Technology Federal Learning team will further carry out technical cooperation with Intel to drive the safe operation and efficient transformation of data resources in federal learning with more and more advanced technologies. Wang Jianzong hopes that the subsequent hardware encryption and decryption environment will be transformed into federal intelligent services. at the same time, Intel can create ecology through industry standard channels in terms of information transmission standards and knowledge training interface specifications. Promote the rapid development and application of federal learning in various industries.

Intel may also develop a training framework to support federated learning, while its related storage technologies, such as SSD (solid state drive), will also develop corresponding industry solutions to further deepen federated learning solutions.

At the network layer, Wang Jianzong believes that the arrival of 5G technology will provide a good opportunity for federation learning, such as solving the bottleneck of parameter exchange to a certain extent, which requires customizing the corresponding technical specifications for federation intelligence at the network communication layer, coding layer, and storage layer. He also said that Ping an Technology is doing research work on relevant federal chips and is also considering whether to design a federal operating system in the future.

However, in order to implement relevant applications and systematically realize the federal intelligence ecology, we have to rely on more frontline artificial intelligence practitioners, who hope that this set of federal learning solution technology can support more companies and industries. carry on the in-depth exploration to the technology based on federal learning, and do some real landing application research.

At present, federated learning technology is more used in the AI training process, and its goal is to form a federated ecology, but Wang Jianzong prefers to achieve federated intelligence through federated database, federated data center, and federated visualization based on federated learning technology. He firmly believes that a new round of innovation derived from new technologies and new requirements will help the AI industry to take off, and federal intelligence is undoubtedly the new dawn of AI innovation.

Https://www.toutiao.com/a6741933088483312140/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.