Introduction _ _
From May 26th to 28th, at the 2021 China International Big Data Industry Expo , with its grasp of cutting-edge trends and technological leadership, the Angel PowerFL secure joint computing technology on the Tencent Big Data-Tiangong platform was awarded the "Leading Scientific and Technological Achievement". Award - New Technology" Award.
The Data Expo is the world's first big data-themed expo, co-hosted by the National Development and Reform Commission, the Ministry of Industry and Information Technology, the State Internet Information Office and the Guizhou Provincial People's Government. As the "highlight" of the Expo, the "Leading Science and Technology Achievement Award" is the only award in the name of the Expo and the only professional award with the theme of big data registered by the National Science and Technology Awards Office.
As the first product of Tencent's self-developed fourth-generation digital-intelligence fusion computing platform "Tencent Big Data-Tiangong", the Angel PowerFL secure joint computing platform focuses on the cutting-edge technology field of privacy-preserving computing. Realize multi-party data collaboration in a secure way, solve the current widespread problem of data silos, and further release the value of data. At present, based on Angel Power FL, Tencent Cloud, Tencent Security, Aegis and other products are opening Tencent's big data's superior privacy computing capabilities to users.
2021 Leading Scientific and Technological Achievement Award - "New Technology": Tencent AngelPowerFL Secure Joint Computing Platform
Cheng Yong, head of Tencent Big Data Angel PowerFL platform, said: "It is a great honor to be recognized by the Data Expo. Angel PowerFL has been investing in research and development since 2019, deeply cultivating core technologies, and has made certain breakthroughs in platform security and performance. While protecting data privacy, it helps data collaboration. It has been implemented in business scenarios such as finance, advertising, and government affairs, and has achieved good application results.”
Taking the financial risk control scenario as an example, a large financial service institution builds a credit risk control model based on the Angel PowerFL platform in conjunction with multi-party data sources when the original data is not local. Compared with the traditional one-sided data modeling, the federated model based on Angel PowerFL has a 5% and 16% improvement in the AUC and KS indicators respectively (see Figure 1 below) - AUC and KS are used to determine the accuracy of the model Commonly used indicators, the larger the value, the higher the accuracy. While improving the accuracy of the model, it also meets the more stringent data privacy requirements in financial scenarios.
Credit Risk Control Model-Comparison of Unilateral Modeling and Federal Modeling
In addition, in the consumer finance business scenario of a bank , federated learning is also performed through the Angel PowerFL platform , and a risk control model is jointly constructed using multi-party data. In the PoC stage, when 1 million test data is used , the KS indicator increases from 0.4 in unilateral data modeling to 0.6 , a 50 % increase (as shown in the figure below), and the effect of federated modeling is remarkable.
In the video recommendation scenario, traditionally, since there is no historical browsing or playback record when a new user of a video APP logs in for the first time, it is impossible to generate a user portrait, and it is impossible to achieve accurate video recommendation for it. Based on the Angel PowerFL platform, using the user's sample label data stored in the other party (or user portrait library), it can help the APP realize intelligent recommendation for new users, and help solve the "cold start" problem of new users. At the same time, user data does not go locally, which fully guarantees privacy.
As shown in the figure below, in the actual scenario, the Angel PowerFL platform helped the video APP's new user retention rate to increase by 21.82%, and the per capita playing time of new users increased by 3.47%, and effectively solved the new user recall problem, bringing significant cost optimization.
Federated Modeling Effects in Video Recommendation Scenarios
In the above three practical scenarios, based on the Angel PowerFL secure joint computing platform, by building a multi-party data joint computing capability to ensure data privacy and security, on the one hand, the model effect is greatly improved on the existing basis; on the other hand, it also By breaking the data silos and realizing data collaboration, it brings new room for improvement to the business system.
With the application and development of new-generation information technologies such as cloud computing, big data and artificial intelligence, mining value from massive data and stimulating the vitality of the data element market have become the main goals of the digital economy era.
Therefore, for data users, it is important to balance the relationship between data fusion and privacy protection, which can not only achieve deeper value mining through data collaboration, but also fully ensure data privacy security, meet compliance requirements, and improve Vibrating the confidence of data owners - finally realizing the "cooperative game" between data fusion and data privacy, and accelerating the development of the digital economy has become the core challenge faced by all enterprises.
Privacy computing has paved the way for this vision to emerge, making it one of the hottest areas of technology in the industry right now. As a multi-field, multi-disciplinary technical system, privacy computing includes three main technical paths, including federated learning, secure multi-party computing, and trusted computing, and involves many technical fields such as artificial intelligence, data science, cryptography, and security protection technology. .
Angel PowerFL secure joint computing platform can provide users with leading privacy computing capabilities. First of all, based on the decentralized architecture, it can effectively avoid single-point security risks. The 3072-bit high-strength homomorphic encryption calculation far exceeds the 2048-bit encryption level requirements of financial supervision; The operation has been optimized at the instruction level, and the performance of Angel PowerFL exceeds the industry by 5 times. The distributed architecture adopted by the platform supports computing on a scale of 100 billion; in addition, it supports joint data modeling and joint data analysis functions, providing end-to-end full scenarios. Private computing capabilities and one-click cloud-native deployment capabilities.
Angel PowerFL technical architecture diagram
Based on the R&D experience and technical accumulation in the fields of big data, machine learning, distribution, security encryption and other technologies, Tencent Big Data Angel PowerFL team has been widely recognized by the industry and won many awards.
Including in the top international privacy computing competition iDASH 2020, the Angel PowerFL team won the championship with excellent results, and its technical strength has been fully confirmed.
iDASH 2020 Award Certificate
In addition, the Angel PowerFL team has also actively participated in the formulation of a number of privacy computing industry-related standards, published many papers at world-class academic conferences, and submitted nearly 20 related technology patents.
In the future, the Angel PowerFL team will continue to forge ahead, work hard in the field of privacy computing, improve platform capabilities, provide assistance for data collaboration in more business scenarios, and release the value of data fusion in a safe and compliant way.