Joint Singapore OpenMined & SIGKDD Meetup - ML, Privacy & Explainability

This is the second Singapore OpenMined meetup and we are organising the event together with the Singapore ACM SIGKDD Chapter.

Agenda:

6.45 - 7.00pm Networking, Food and Drinks
7.00 - 8.30pm Talks by Assist Prof Reza Shokri (NUS Presidential Young Professor of Computer Science) and Naresh Rajendra Shah (Co-founder and CTO at Untangle AI)
8.30 - 9.00pm Discussion & Networking

The event is supported by SMU School of Information Systems.

### Talk 1 Details ###

Speaker:
Assist Prof Reza Shokri (NUS Presidential Young Professor of Computer Science)

Title:
Data Privacy in Machine Learning: from Centralized Platforms to Federated Learning

Abstract:
In this talk, I will give a broad overview of data privacy risks in machine learning systems. I will show how an adversary can exploit the privacy vulnerabilities of machine learning algorithms using inference attacks. I will then present privacy-enhancing algorithms that can limit the information leakage about sensitive data, while enabling meaningful computations. Examples of these algorithms include differential privacy, trusted hardware, secure multi-party computation, and federated learning.

### Talk 2 Details ###

Speaker:
Naresh Rajendra Shah (Co-founder and CTO at Untangle AI)

Title:
Tests and metrics to evaluate ML model explanations

Abstract:
With a variety of explanation methods available today, how do we understand the limitations and pitfalls of the explanation methods? Explanation methods are by themselves ML problems and in that view, to make them robust, we bring along a test suite similar to metrics and tests available for most ML methods today. If you have a new explainability method, then you can test that method against these tests. If you want to use an explainability method, you will be aware of its limitations and know when is the problem arising from the explainability method as opposed to the model itself. Lastly, this allows for incremental progress as well as in some cases defence against adversarial attacks of specific kinds.