Textbook in PDF format

In this book, the authors highlight the latest research findings on the security and privacy of Federated Learning systems. The main attacks and counterattacks in this booming field are presented to readers in connection with inference, poisoning, generative adversarial networks, differential privacy, secure multi-party computation, homomorphic encryption, and shuffle, respectively.
The book offers an essential overview for researchers who are new to the field, while also equipping them to explore this “uncharted territory.” For each topic, the authors first present the key concepts, followed by the most important issues and solutions, with appropriate references for further reading.
In the recent two decades, we have witnessed the dramatic development of Artificial Intelligence (AI in short), not only in Artificial Intelligence itself but also its applications in various sectors of human society. The appearance of deep learning pushed AI into another spring after a long winter. Nowadays, there are many successful stories of significant progress in various scientific fields with the help of AI, for example, the application of AI in biology, chemistry, law, and social science, to name a few.
However, Artificial Intelligence (AI) suffers a fundamental challenge, explainability: the conclusions obtained from Machine Learning based on big datasets are generally very useful, but we sometimes do not know why. People even treated AI as alchemy: the outputs of the same AI algorithm with the same inputs may vary from time to time, and AI practitioners sometimes even do not know what they will have before the deployment of AI.
Security and privacy protection in AI is far behind the fast development of AI and its applications. As we can see, AI is gradually permeating into our daily lives, and security and privacy are big challenges as AI needs sufficient information for their judgment and recommendation. As a result, we can see that AI and privacy protection are natural enemies. We can predict that majority (if not all) known attacks will be applied in AI applications, we need solutions. Particularly, digital privacy is an unprecedented challenge, and we face numerous new problems, for example, measurement of privacy, privacy modelling, privacy tools, and privacy pricing, to list a few.
In general, Federated Learning (FL) is a branch of Deep Learning, which is a powerful tool to address various complex problems in the past decades. Google proposed Federated Learning as a variation of Deep Learning in order to address the privacy concern from data owners. Federated Learning (FL) is a big step for privacy protection in Machine Learning; however, it is not perfect. The Federated Learning framework allows learning participants to keep their data locally and download the training model from a central server or servers to execute a local training. The updates will be uploaded to the server(s) for a further aggregation for the next round of training until an acceptable global model is reached. Despite the advancement of its computing model, Federated Learning still faces many security and privacy challenges, which have attracted a lot of attention from academia and industry.
We classify the security and privacy research in Federated Learning into two categories: problem based and tool based. In our understanding, problem-oriented research focuses on problems and proposes solutions. At the same time, tools-oriented research usually depends on tools to address problems.
1 Introduction to Federated Learning
2 Inference Attacks and Counterattacks in Federated Learning
3 Poisoning Attacks and Counterattacks in Federated Learning
4 GAN Attacks and Counterattacks in Federated Learning
5 Differential Privacy in Federated Learning
6 Secure Multi-party Computation in Federated Learning
7 Secure Data Aggregation in Federated Learning
8 Anonymous Communication and Shuffle Model in Federated Learning
9 The Future Work