Introduction to Differential Privacy (Cambridge machine learning reading group)
Date:
With Ferenc Huszár. Slides (first half only). Talk.
Privacy is increasingly important to society and coming under more scrutiny. Modern machine learning methods consume vast amounts of data, and it is well known that many large networks memorise their training data. It is necessary to move beyond ineffective heuristics such as anonymisation or selective withholding of data. Differential privacy provides a framework for quantifying loss of privacy and designing algorithms which keep reasonable worst case privacy loss within acceptable levels. In this talk, we will motivate and introduce foundational differential privacy methods, and look at applications to machine learning.
Recommended reading: Chapter 1 of The Algorithmic Foundations of Differential Privacy; Cynthia Dwork and Aaron Roth (2014) (link)
Optional reading: Deep Learning with Differential Privacy; Abadi et al (2016) (link)