[Pearg] differential privacy in dynamic systems

[Amelia Andersdotter <[e-mail]>]

Hi all,

To kick off a differential privacy discussion, I would like to share a
recent find:
http://sist.shanghaitech.edu.cn/faculty/luoxl/class/2017Fall_EE251/GlobalSIP2017/pdfs/0000487.pdf


It covers differentially private Kalman filtering, or differential
privacy in dynamic environments. I send it because the layout of what
differential privacy *does* is fairly straight forward: the differential
mechanism adds noise to data released after a query, which in the case
of a dynamic measurement necessarily implies reducing the quality of the
released data.

In the Kalman filter, this is particularly obvious since measurements
are already assumed to be noisy. Adding more noise makes measurements
more noisy. The paper creates a differentially private estimator by
organising the "real" measurements in blocks, and then adding noise to them.

I think the structure of the paper also displays some of the big
research topics in differential privacy:

- existence of the transformation matrix D, which organises measurements
in blocks.

- computational tractability of transformation matrix D. This is
typically a tricky problem.

- estimation of how much "worse" estimators get by making measurements
more noisy (cf. the final example in the paper).

In relation to the last point, this paper by John Duchi and Michael I.
Jordan covers the loss of accuracy in estimators from differentially
private released data: https://arxiv.org/abs/1604.02390

There are a bunch of really interesting questions that arise from the
above presentations: 1) when can we deal with estimators being worse? 2)
the data aggregator will continue to have full control over the entire
dataset that is being queried: differential privacy is
aggregator-centered, because it assumes the aggregator will process the
data before handing it out, 3) should measurements be made more noisy at
collection, rather than at query and how does that impact power
relations between aggregators and queriers under question 2).

best regards,

-- 
Amelia Andersdotter
Technical Consultant, Digital Programme

ARTICLE19
www.article19.org

PGP: 3D5D B6CA B852 B988 055A 6A6F FEF1 C294 B4E8 0B55