Today's artificial intelligence is still facing two major challenges. One is, in most industries, the data in the form of an island. Second is to strengthen data privacy and security. We propose possible solutions to address these challenges: security federal study. In addition to federal learning framework Google first proposed in 2016, we also introduced a comprehensive security federal learning framework, including the federal level learning, learning vertical federal and federal transfer learning. We offer a definition, architecture, and applications for the federal learning framework, and provides a comprehensive survey on the subject of existing work. In addition, we propose to establish mechanisms between organizations based on federal data network, as an effective solution to allow the sharing of knowledge without compromising user privacy.
1 INTRODUCTION
2016 artificial intelligence (AI) maturity of one year. With AlphaGo [59] Go beat the top human players, we really witnessed the tremendous potential of artificial intelligence (AI) and many come to expect, including unmanned vehicles, health care, finance and other applications, the use of more complex , sophisticated artificial intelligence techniques. Today, artificial intelligence technology can play to their strengths in almost all industries. However, reviewing the development of artificial intelligence, the development of artificial intelligence inevitable experienced several ups and downs. The AI will do a turn? When will appear? Because what factors? Artificial intelligence part of the current public interest is driven by the large data availability: 2016, AlphaGo use a total of 300,000 games as training data to obtain excellent results.
With the success of AlphaGo, people naturally want to be like AlphaGo such a large data-driven artificial intelligence can be achieved quickly in every aspect of our lives. However, the reality is somewhat disappointing: in addition to a few industries, most areas of the data are limited or of poor quality, so as to achieve artificial intelligence more difficult than we thought. Is it possible to transfer data across the organization, the integration of data into a common site? In fact, in many cases, break down barriers between the data source, if not impossible, is very difficult. Generally speaking, any data needed artificial intelligence project involves a variety of types. For example, artificial intelligence-driven product recommendation service, the seller has product information, user purchase data, but does not describe the user data capacity purchase and payment habits. In most industries, the data in the form of an island. Since the data between the different sectors of industry competition, privacy, security and management of complex procedures, even the same integrated company is also facing great resistance. Almost impossible to disperse in data and institutions across the country to integrate, otherwise the cost is unbearable.
At the same time, with the big companies in the sense of compromise data security and privacy of users increasing, the importance of data privacy and security has become a major problem worldwide. News about open data leakage caused great concern to the public media and the government. For example, recent data breaches Facebook sparked widespread protests [70]. In response, the world is legal protection to enhance data security and privacy. For example, the European Union on May 25, 2018 implementation of the "General Data Protection Regulation" (GDPR) [19]. GDPR (Figure 1) designed to protect the user's privacy and data security. It requires companies to use clear language of the User Agreement and grant the user the "forgotten right", that is, the user can delete or withdraw their personal data. The company in violation of the bill will face stiff fines. The United States and China are also developing similar privacy and security behavior. For example, in 2017 promulgated the "China Security Law" and "Civil Law" requires Internet companies may not disclose or tamper with their personal information collected during data transactions with third parties, we must ensure that the proposed contract follows the legal data protection obligations. Establishment of these regulations will significantly contribute to a more civilized society, but will also create new challenges in artificial intelligence, common data processing program.
More specifically, artificial intelligence traditional data processing model typically involves simple data model transaction, one party to collect data and transfer it to another party, the other party responsible for clean-up and integration of data. Finally, the third party will acquire and build an integrated data model for others to use. These models are usually sold as a final product of service. The traditional procedure faced with the challenge of these new data regulations and laws. In addition, because users may not know the future use of the model, these transactions violated GDPR and other laws. Therefore, we are faced with a dilemma: our data is in the form of an island, but in many cases, we are prohibited collection, integration and use of data to different places artificial intelligence processing. How to legally solve the problem of data fragmentation and isolation are the main challenges of today's artificial intelligence researchers and practitioners face.