sana.txt

Machine learning is widely used in practice to produce predictive models for applications in business, health care, 
recommendation services, threat analysis, and authentication technologies. These models are more accurate when trained on huge
amount of data collected from different sources. However, the massive data collection raises privacy concerns. An ML model may
inadvertently and implicitly store some of its training data; therefore, careful analysis of the model may reveal sensitive
information. Selected literature deals with the problem of distributed and privacy preserving machine learning (ML) and
provides different approaches to address the problem. The first paper demonstrates a generally applicable multi-tier ML
approach to providing strong privacy guarantees for training data and proposes a distributed learning algorithm. The second
paper presents a system for scalable privacy-preserving ML using a multi-party computation (MPC) approach. The third paper
offers another interesting approach, privacy-preserving ML as a service, where they leverage the use of private computation
units (Intel SGX) to reveal neither the training algorithm nor the model structure to the user, hence providing only black-box
access to the trained model.