Differential privacy

From Simson Garfinkel
Revision as of 05:19, 29 October 2020 by Simson (talk | contribs) (→‎Videos)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search

A few references on Differential Privacy, for people who don't want to get bogged down with the math.

View Differential Privacy on Google Trends

And the US Census Bureau

There are now official decision memos! Two, dated July 1, 2019 cover the only official statements on the design and application of Differential Privacy to the 2020 Census that have been made to date.

The memo https://www2.census.gov/programs-surveys/decennial/2020/program-management/memo-series/2020-memo-2019_13.pdf states the Group Quarters invariant, which is "number and type of GQ facilities."


Text Materials



Database Reconstruction

The idea of that releasing multiple queries on a confidential database could result in the reconstruction of the confidential database goes back to the 1970s.

We explain how to perform database reconstruction in our 2018 ACM Queue article:

This article summarizes the risks of database reconstruction, as understood in 1989:

I learned of the connection from Dorothy Denning's work on The Tracker:

Dinur and Nissim's "Database Reconstruction Theory" is actually a proof that random queries on a database, which can be generated with complexity P, will reveal the full contents of the database:

But query auditing was shown to be NP-hard in 2000:

  • J. M. Kleinberg, C. H. Papadimitriou and P. Raghavan, Auditing Boolean Attributes, PODS 2000

So the only way to protect against a large number of unaudited queries is to add noise to the database. The proof in Dinur and Nissim is that adding noise protects against *all* queries, random and otherwise. The more noise, the more protection.


Foundational Papers

Critical Papers


Public Perception


Existing Applications

On The Map, at the US Census Bureau

RAPPOR, in Google Chrome



Advanced Topics

Differential Privacy and Floating Point Accuracy

Floating point math is not continuous, and differential privacy implementations that assume it is may experience a variety of errors that result in privacy loss. A discussion of the problems inherently in floating-point arithmetic can be found in Oracle's What Every Computer Scientist Should Know About Floating-Point Arithmetic, an edited reprint of the paper What Every Computer Scientist Should Know About Floating-Point Arithmetic, by David Goldberg, published in the March, 1991 issue of Computing Surveys.

"How Will Statistical Agencies Operate When All Data Are Private?" (MS #1142) has been published to Journal of Privacy and Confidentiality. http://repository.cmu.edu/jpc/vol7/iss3/1

The Fool's Gold Controversy

Other attacks


p for randomized response rate:

$p = \frac{e^\epsilon}{1+e^\epsilon}$

Probability that randomized response should be flipped.

See Also