4 ms·
I was curious how this application approached privacy budget management (e.g. how the privacy parameter ε is accounted for over multiple searches), but, flippin
by sirgawain33 7y ago
I was curious how this application approached privacy budget management (e.g. how the privacy parameter ε is accounted for over multiple searches), but, flipping through the source, this application doesn't appear to use Differential Privacy at all.
The anonymization approach implemented is "generalization". Here's a test showing the outputs this app would produce:
https://github.com/abe-winter/arbout/blob/master/test/test_diff_summary.py#L40 https://github.com/abe-winter/arbout/blob/master/test/test_d...
- awinter-py 7y agoyou're right, edited the title
- awinter-py 7y agono wait, I'm not actually sure you're right differential privacy is any search that protects individual inputs from disclosure
- sirgawain33 7y agoDifferential Privacy is a mathematical property. Here is the precise definition: https://en.wikipedia.org/wiki/Differential_privacy#Definition_of_%CE%B5-differential_privacy https://en.wikipedia.org/wiki/Differential_privacy#Definitio... One way to phrase it intuitively is "the probability of any particular output from two databases that differ by one element is almost the same". The bound on "almost" is captured by the privacy parameter ε. One of the smoking guns that this algorithm is not differentially private is that the code doesn't import `random` anywhere! A differentially private algorithm is always going to be stochastic.
- awinter-py 7y agoI've removed refs to differential privacy from the repo -- I think you're right and I misused the term