3 ms·
Each year, I put up lists of interesting papers I've read across CS, math, econ, etc. Here's the list from 2022: https://bcmullins.github.io/interesting-article
by wannabebarista 3y ago
Each year, I put up lists of interesting papers I've read across CS, math, econ, etc. Here's the list from 2022: https://bcmullins.github.io/interesting-articles-2022/ https://bcmullins.github.io/interesting-articles-2022/.
That should give you dozens of recommendations and a bit of rationale for them.
It's hard to say what best means, but some papers play their part better than others. Not every paper should change the field, nor should they all be the length of a novel.
- tnecniv 3y agoDo you have more recommended reading on differential privacy from recent years?
- wannabebarista 3y agoHere are a few papers on some scattered DP topics. I'm working on a guide to learning DP now similar to one I've written on measure theory in the past. The Matrix Mechanism (2015) https://people.cs.umass.edu/~miklau/assets/pubs/dp/Li15matrix.pdf https://people.cs.umass.edu/~miklau/assets/pubs/dp/Li15matri... Good introduction to the select-measure-reconstruct approach to answering large sets of linear queries (e.g. histograms, averages, ...). This paper is important for the statistics approach to DP synthetic data (which is what I primarily work on): https://differentialprivacy.org/synth-data-1/ https://differentialprivacy.org/synth-data-1/ Differentially Private Quantiles (2021) https://arxiv.org/pdf/2102.08244.pdf https://arxiv.org/pdf/2102.08244.pdf It turns out that medians are difficult to deal with under DP. This paper introduces a method called JointExp to estimate several quantiles at once. One innovation here is that the user doesn't have to know anything about the distribution of the variable in advance for it to work. Boundary Differential Privacy (2020) https://www.computer.org/csdl/journal/tq/5555/01/09286504/1porkN6i0IE https://www.computer.org/csdl/journal/tq/5555/01/09286504/1p... The main motivation for DP is to protect the privacy of people whose data is used to compute some statistic. This paper, instead, uses DP techniques as a way to inject controlled noise into the output distribution of a classifier to defend against model reconstruction attacks.
- tnecniv 3y agoThanks! I know some basics about DP and have adapted some of their results for my non-privacy related work, but it’s not an area I know much about the cutting edge of. Your blog post inspired me to catch up on some results from the past few years!
- wannabebarista 3y agoNice! If you ever want to chat about DP stuff, feel free to shoot me an email.