Recent Advances in Differential Privacy under Continual Observation

Differential privacy is one of the most popular definitions of privacy, being used both in research as well as in industrial applications. The multi-query setting, where many queries have to be answered over a static data set, is already well understood for many algorithmic questions. The dynamic setting, where updates to the data set are interleaved with queries over the current data set, was introduced in 2010 by Dwork, Naor, Pitassi, and Rothblum, who called it differential privacy under continual observation. While there was not much work on that model in the 2010s, it has received significant attention in recent years, partially due to its use in private machine learning. I will survey the state-of-the-art in the continual observation setting and explain its motivation as well as the main algorithmic techniques that have led to the recent advances.