Joel Daniel Andersson

I do research on differentially private algorithms as a member of the Providentia group under the supervision of Rasmus Pagh. In particular I study such algorithms in the continual observation setting. The setting can be summarised as follows:

1. There is an input stream S = {x_i : i ≥ 1} where x_t is received at time t ≥ 1.
2. At each time t, produce an output y_t that is a function of S_1:t={x_i : 1 ≤ i ≤ t}.
3. We want the output sequence {y_i : i ≥ 1} to satisfy differential privacy for neighbouring streams S, S' where the streams differ in exactly one element at some time t=j where ||x_j-x_j'||_p = 1, p typically being 1, 2 or ∞.

One of the simplest problem in this setting is binary continual counting where x_i ∈ {0,1} and we want to output the number of 1s in the stream up to each time t, i.e. the prefix sum up to each time t. Without the constraint of differential privacy the problem is trivial, but with privacy included we do not know how to solve the problem optimally for any DP variant.

Continual observation is also inherently practical: releasing aggregate statistics over time is useful. If simultaneously privacy is a consideration, then continual observation is a good framework. Such a situation might for example be continually releasing disease statistics. And if that does not suffice as motivation, then it can also be added that algorithms for continual observation are used as subroutines in differentially private machine learning.

That said, you can also catch me on https://jdandersson.net/.