Ant Financial has reshaped payments and changed lives over the past 15 years, providing services to more than 1.2 billion people around the world, which is supported by technology. At the 2019 Hangzhou Computing Conference, Ant Financial will share its technological precipitation over the past 15 years, as well as its future-oriented financial technology innovation with attendees. We have compiled some of the best speeches and will continue to publish them in the”
Ant Financial Technology“On the public account, this article is one of them.

In the era of artificial intelligence, data is the oil in the FIELD of AI. Without data, it is difficult to better implement AI. However, data islands hinder the acquisition and utilization of data. Ant Financial started to deploy privacyprotection machine learning three years ago, committed to machine learning under the premise of data security and privacy protection, we call it shared intelligence. We’ve talked about the concept and principles of shared intelligence before. Today, we want to talk about the development and application trends of shared intelligence.

The problem with AI is that it can’t have it both ways — privacy and usability. If you want your AI system to work, you may need to sacrifice privacy. However, in a large number of real-world scenarios, failure to balance privacy and usability can cause a lot of DIFFICULTIES for AI landing.

Just a few examples.

First loan risk control, and user want to buy A house to bank loans, the bank may be judged to be A “bad”, there is no way to give him on loan, because the institutions holding the one part of the data, the same user B to institutions, the institution B based on the partial data it has, it is possible for his loan, this contradiction is everywhere, All due to data inconsistency.

In the field of medical wisdom, some rare diseases in each hospital’s case is not much, if we can keep the case sharing between hospitals, can get more sample data, so as to make more accurate diagnosis can make use of AI, but technology is not the priority in the inside of the case, for the hospital, it has the responsibility to protect the privacy of patients, how to ensure that in a Shared case at the same time, Not giving away users’ privacy is the first priority.

The problem of data silos presents many similar challenges for AI landing and application.

In the real world, the data in this diagram is not connected, there may be some temporary links somewhere, most of the data in this diagram is disconnected. Our goal is to break through data silos and use technology to solve technology problems. Data sharing and value transfer can be realized while data security is protected by technology.

Shared intelligence: Available not visible

For shared intelligence, we hope to achieve the goal of invisible data availability. In the scenario of multi-party participation and mutual distrust between data providers and platform parties, multi-party information can be aggregated for machine learning, and the privacy of each participant will not be leaked and the data will not be abused.

To achieve this goal, we use many existing technologies in the industry, such as differential privacy, which has been studied in academic circles, trusted execution environment, which has been explored by many big data vendors, and multi-party secure computing, which has gained wide attention with the improvement of computing power and hardware technology + cryptography breakthrough. In some cases, there is less target data but more source data, so we use the method of transfer learning to share data, which also belongs to our big technology category.

Concrete, the first kind of plan is the trusted execution environment, mainly depends on the middle of the hardware level safe Enclave, the two sides through some mechanism of cryptography, the data is encrypted, only inside the suitcase card can decrypt after encryption, decryption since a wide variety of computing, the new type of suitcase card because the suitcase card is the trusted third party, We do not trust each other, trust the password box, so that in the case of data privacy will not be disclosed, to do all kinds of AI algorithms.

This scheme relies on trusted hardware, encrypts data, and centrally transmits data to trusted platforms. For some organizations, they already have access to the cloud, so that all things are stored in the cloud and all technologies are deployed in the cloud, so it is very fast and convenient to adopt this method, and at the same time can achieve a good effect of privacy protection.

The second scheme is a software-level scheme, in which we do the corresponding data processing before calculation. For example, the secret sharing technology, by splitting the data, several parties send random numbers to complete the calculation, and then can complete all kinds of AI calculations and models; There are also methods like homomorphic encryption, in which corresponding operations are performed in the encrypted space to complete the computation of AI, and there is a control module in the middle to jointly complete the learning goal. This method itself does not involve hardware, is a partial software + cryptography scheme, the middle out is random number/encryption intermediate results, currently in the industry privacy +AI combination direction, this scheme is relatively more.

Nebula: A Shared Intelligent Network

Shared intelligence requires a lot of involvement, and we’ve designed Nebula shared Intelligence Network architecture. For Ant Financial, we want to create this shared intelligence network with a lot of partners.

There are various computing nodes in the network, which can trigger AI calculation in a management platform. This shared intelligent network, can use different technologies to achieve the goal of shared intelligence, such as the construction of joint marketing network, between nodes can be arbitrary networking, multi-party security computing technology to achieve joint marketing, while the management node can be deployed anywhere; For some organizations, they may not have strong AI capabilities and multi-party computing capabilities, so they can rely on cloud technology, put data in a trusted execution environment, participate in the construction of such a network, and solve the problem of the last kilometer of AI landing through such shared intelligent technology.

The architecture of our entire compute node is shown in the figure above. The bottom layer is similar to the normal environment, with various trusted execution environments on the left and normal CPU and GPU environments on the right. There will be a uniform API layer to mask these different details.

Further up, there will be native computing, which itself will be slightly different from the common open source frameworks, and we will move from the current popular version to a secure version, such as secure XGBoost. When MPC is made in the middle, we will provide various technologies, such as obtrusion circuit, OT and so on. The top layer provides some visual and interactive interfaces, through which ordinary users can complete complex multi-party calculation operations. It also supports various security model inference to protect privacy.

We want to share intelligence technology through this architecture, and create a visual interface, using a drag and drop approach to complete the construction of AI computing quickly and efficiently.

The shared intelligence architecture mentioned above has now achieved the goal of better completeness, ease of use and stability, and has been implemented in many places. In terms of completeness, we have achieved functional completeness and scene completeness. At present, we mainly support risk control and other typical AI scenarios. The algorithms in the system are relatively comprehensive, covering linear model, tree model, deep learning, graph neural network and other directions. In terms of ease of use, we hope to promote this modeling technology, and at the same time “shield” some underlying technologies (trusted execution environment, multi-party secure computing, etc.) to reduce the cost of learning and using. In terms of stability, we realize the clustering of shared intelligent computing and support remote operation and maintenance.

We have put shared intelligence online on the big data intelligent platform. The following demo is an AI modeling demonstration of multi-party secure computing.

The previous pre-processing part looks the same as normal AI modeling. After the data is pre-processed through drag and drop operation, it is sent to the shared intelligent modeling, which will produce the results of AI calculation. In this way, the threshold for using new technologies can be greatly reduced, facilitating the use of business parties.

Ant Financial has been building in the field of shared intelligence for more than three years, published more than 10 papers and obtained more than 80 patents. In terms of standard project approval, we are simultaneously promoting IEEE Shared intelligence, ITU-T MPC international standard, CCSA Shared intelligence industry standard and AIOSS/AIIA Shared Intelligence Alliance standard. It has also won several awards for innovation.

Share intelligent landing cases

Next, share three typical landing cases.

One is in the field of security risk control, to establish security risk control network with ecological partners. Ecological partners use the trusted execution environment technology introduced above to encrypt data and transmit it to the network to build this model, crack down on fake transactions and gang crimes, greatly improve the accuracy of risk control and realize the purification of risk control network. Through such a risk control network platform, businesses can add a lot of transactions every day, while reducing capital losses.

The second one is Zhonghe Rural Credit. We greatly improved risk control performance through data fusion, and changed the traditional offline mode into the online automatic review mode. It only took 5 minutes to complete credit granting, with a total loan of 3.19 billion yuan and 440,000 people successfully granted credit in 8 months. We will help achieve inclusive finance in rural areas.

The third is the joint credit risk control with The Bank of Jiangsu. Remember our previous example? Incomplete data leads to wrong risk control decisions. Now, by sharing intelligent technology, both parties can complete the common model construction and realize joint risk control through such a mechanism, which greatly improves the effect. In this process, users’ data and privacy are effectively protected.

In general, we want to build an open and shared intelligent network. We hope that more partners and institutions will join us to complete the construction, break data silos, and facilitate the implementation and application of AI technology.

OceanBase topped the TPC-C test list and achieved the breakthrough of zero database in China. Want to know the technical details behind? Welcome to download the e-book “OceanBase TPC-C Test Technology Analysis”. Long press to identify the following QR code, follow the official official account of “Ant Financial Technology”, and reply “TPCC” in the dialog box, you can download it for free.