samuel-maddock / pure-ldp Goto Github PK

Recent work on empirical LDP auditing [1], has highlighted a small bug in the _perturb() implementation of UEClient where the users true value was not properly retained with probability $p$. This results in UEClient perturbation failing to guarantee $\varepsilon$-LDP, particularly for small values of $\varepsilon < 1$ (See Figure 3 in [1]).

As of fc622e3, this has been fixed and will be rolled out in version 1.2.0. It is recommended to update the package if you are currently using UEClient/UEServer.

[1] "Revealing the True Cost of Local Privacy: An Auditing Perspective", Héber H. Arcolezi and Sébastien Gambs, https://arxiv.org/abs/2309.01597

Hi @Samuel-Maddock ,

I seem to experiencing some issues when trying to run the code in 'Usage'. It throws an error originating from hr_server.py in the pure-LDP/pure_ldp/frequency_oracles/hadamard_response/ saying that the function k2k_hadamard couldn't be found. Looking at the import, it doesn't seem like it could find it because I don't see an internal folder in hadamard_response/ either.

Here is the import that is causing the error: from pure_ldp.frequency_oracles.hadamard_response.internal import k2k_hadamard

Best,
Ömer

I am currently trying some things out where my array has a lot of entries (more than a million). I think in the example that you @Samuel-Maddock used, you had around 18,000 entries in your data array. Any suggestion on how to speed the for loop up?

Is there a version with the IRR also implemented for the rappor clients?