Practical Secure Inference | Asia Innovation Summit

Fueled by massive data and the availability of extensive compute, sophisticated machine learning models have found diverse applications across verticals such as healthcare and finance. This has made the problem of privacy-preserving machine learning increasingly important. In this talk, I focus on \secure prediction-as-a-service. Here, a hospital/ML provider has a machine learning model that has been trained on sensitive data, and patients have their private medical records. The goal is to enable the patients to learn the prognosis based on the model without revealing their sensitive medical data while preserving the confidentiality of the model held by the hospital. Our work CrypTFlow provides a programmable, scalable and efficient cryptographic solution for secure inference. CrypTFlow is a system that automatically compiles TensorFlow/ONNX inference code to secure computation protocols. It has two components. The first component is an end-to-end compiler from TensorFlow/ONNX to a variety of secure computation protocols. Second, we build specialized protocols for secure machine learning for two and three-party settings that are orders of magnitude more performant than prior works. I demonstrate our results on multiple case studies from the healthcare domain.

 

Date:
Haut-parleurs:
Divya Gupta
Affiliation:
Microsoft Research India