Privacy Preserving Machine Learning

In the light of huge interest in using machine learning algorithms and simultaneous requirement of security of data, the field of privacy-preserving machine learning (PPML) has emerged as a flourishing research area. Typical supervised learning consists of two phases: (1) the training phase during which the algorithm learns a model “w" from a data set of labelled examples, and (2) the classification phase that runs a classifier “C" over a previously unseen feature vector “x", using the model to output a prediction “C(x,w)". A privacy-preserving classification protocol facilitates interaction between a server (whose input is a model “w") and a client (whose input is a feature vector “x") such that at the end of this interaction the client should learn “C(x,w)" but nothing else about the model “w", while the server should not learn anything about the client's input “x".

The additional demand on privacy makes the already compute-intensive ML algorithms more demanding in terms of high compute power. It is economical and convenient for end-users to outsource an ML task to more powerful and specialized systems. However outsourcing computation to server must be done in a way such that privacy of data is not violated. A potential solution explored in this area is SOC - secure outsourced computation. SOC allows end-users to securely outsource computation to a set of powerful cloud-servers and avail its service on a pay-per-use basis, while guaranteeing the privacy of the end-user's data against malicious servers. In this project, we propose to work on the following lines:

• To explore PPML techniques in the SOC setting for widely used ML algorithms – Linear Regression, Logistic Regression, and Neural Networks.
• Study and implementation of a few state-of-the-art customized privacy-preserving building blocks, such as dot-product, truncation, comparison, rectified linear unit (ReLU), Sigmoid etc.
• Deeper understanding of the privacy requirements for ML and Federated ML

Researchers : Soham De, Kuber Shahi, Mahavir Jhawar, Subhashis Banerjee