Summary and Outlook
In this book, we gave an overview of federated machine learning (a.k.a. federated learning), mainly covering the background information on privacy-preserving machine learning and distributed machine learning, horizontal federated learning, vertical federated learning, federated transfer learning, incentive mechanisms, federated learning in computer vision, natural language processing and recommender systems, federated reinforcement learning, and applications of federated learning in various industrial sectors.
Federated learning was born out of the increasing concern over data fragmentation, data silos, user privacy leakage, and data shortage problems facing machine learning. Our society is becoming aware of the severe impact of user privacy violations by large corporations, and regulators are tightening the laws governing the sharing of private data, such as the most stringent requirements of the GDPR on data security [DLA Piper, 2019]. As the traditional machine learning approaches that are based on centralized data collection is no longer compliant with strict data protection laws, in order for the field of AI to continue advancing, an innovative solution that can preserve data privacy is badly needed.
Federated learning allows multiple parties to hold their own data privately while building a machine learning model collaboratively and securely. With federated learning, data does not need to leave the data owners, and hence privacy can be better protected. In this book, we have discussed several modes in building the federated machine learning model, including horizontal federated learning, vertical federated learning, and federated transfer learning. We have also explored federated reinforcement learning, federated learning in computer vision, natural language processing and federated recommendation systems. As societies move ahead, these techniques are likely to play a major role in moving AI into the next level, a level in which models can be built collaboratively, confidentially and in a privacy-preserving manner. One such example is the Google Gboard system [Bonawitz and Eichner et al., 2019]. In this book, we have also pointed out that federated learning is not only just a technical solution, but also a privacy-preserving ecosystem, such as the FedAI ecosystem [WeBank FedAI, 2019]. Building a federated learning ecosystem is also an economic problem in which incentive mechanisms need be carefully designed to ensure that the profits are shared fairly and transparently. This is so that different parties will want to join the federation in a sustainable way. In addition, such mechanisms should also help federations dissuade malicious participants.
As more application scenarios for federated learning are being explored, the field is becoming ever more inclusive. It spans research and practice in machine learning, statistics, information security, encryption and model compression, game theory and economic principles, mechanism design, and more.
In the future, the federated learning ecosystem is likely to expand. More open-source software will emerge, such as FATE [WeBank FATE, 2019] and PySyft [2019]. Practitioners will be accustomed to building solutions that have all the necessary facets expected by the society, and federated learning will become a prime example of “AI for Social Good.”