BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//project/author//NONSGML v1.0//EN
CALSCALE:GREGORIAN
BEGIN:VEVENT
DTEND:20221125T120000Z
UID:366236faa1cb65b315bcd3cfacc67430-361
DTSTAMP:19700101T120011Z
DESCRIPTION:Multi-Party Computing for Privacy in Machine Learning Systems
URL;VALUE=URI:https://www.csa.iisc.ac.in/newweb/event/361/multi-party-computing-for-privacy-in-machine-learning-systems/
SUMMARY:Given the resource management benefits such as elasticity, availability, and cost-effectiveness offered by cloud service providers, a growing number of machine learning workloads are migrated to the cloud for operations. Under this modern compute paradigm, confidential data and models can be leaked to unwanted parties if the service providers are curious, malicious, or compromised. The privacy concern is particularly pressing for natural language processing (NLP) where user’s audio features are inputs to ML models. These inputs contain sensitive private information about the users and require rigorous protection. 

Secure multi party computing (MPC) is one approach to tackle the privacy leaks without relying on any additional hardware support. MPC protocols provide strong security even when a subset of parties are compromised. However, when it comes to protecting privacy there is no free lunch, and in fact we show that it is a very expensive lunch. Through a detailed characterization of industry-strength MPC implementation of Transformer-based NLP models, we analyze where the MPC performance bottlenecks are. First, we show that Transformers rely extensively on softmax
<br>
Talk link <a href="https://youtu.be/2nE6nbkfuls">https://youtu.be/2nE6nbkfuls</a>
<br>
DTSTART:20221125T120000Z
END:VEVENT
END:VCALENDAR