Privacy Preservation in Federated Learning: its Attacks and Defenses

In the context of artificial intelligence, the traditional centralized server-based machine learning model training approaches have significant security and privacy issues. The federated learning (FL) approach was created to address these problems. FL uses a privacy-by-design architecture, in which several devices work together to develop any machine learning system that does not reveal users' personal information under the control of a single server. In federated learning, privacy preservation is achieved through several techniques that involve adding noise, secure aggregation, local training, and model compression. These techniques aid in making sure that raw data is not transmitted over the network and that only the required information is shared to train a global model. This research study has discussed about the federated learning-related privacy threats and defenses. The existing research works are divided into two types as attack methods and corresponding defense mechanism and further analyzed the relevant research papers from the last few years, and then further discussed the research challenges, and finally concluded the research study.