Privacy Preservation in Federated Learning
P. Keerthana, M. Kavitha, Jayasudha Subburaj · 2024
Workflows for machine learning require numerous individuals acting in various capacities. For instance, users may interact with their devices to generate training data, which is then used in a machine-learning training procedure to extract cross-population patterns (e.g., trained model parameters). A machine learning engineer or analyst can then evaluate the quality of the trained model, and in the end, the model may be made available to end users to support particular user experiences. A novel approach to preserving user privacy during cooperative model training across dispersed devices is provided by federated learning. Strong safeguards that preserve private data throughout the federated learning process are essential as sensitive data continues to proliferate on individual devices. The main tactics and methods used in federated learning to protect privacy are examined in this chapter. Additionally, it is shown that the idea of model personalization works well, enabling devices to gain from a global standard that is customized to local details without jeopardizing privacy. By guaranteeing the privacy of the initial models that are sent to every device, secure initialization approaches add to the entire privacy framework. The necessity of auditability and monitoring in federated learning systems is emphasized in this chapter in addition to technological measures. To identify and resolve any privacy violations and make sure that privacy-preserving measures continue to work in the face of changing risks, periodic evaluations and assessments are crucial.