Floating-point privacy vulnerabilities

[naoise-h]

Vulnerabilities of the Laplace mechanism to floating-point attacks have been flagged in the past. Fixing this vulnerability and investigating similar vulnerabilities for other mechanisms is highly desirable.

[danrr]

Hi @naoise-h. I implemented the Mironov paper (at https://github.com/danrr/Snapping-mechanism) to get a better understanding of it. The attack code I wrote seems to work against diffprivlib's Laplace mechanism and can differentiate between 0 and 1 with a high probability.

I have implemented the snapping mechanism as described in the paper (https://github.com/danrr/Snapping-mechanism/blob/main/SnappingMechanism/SnappingMechanism.py) so I've started adapting the code to integrate with diffprivlib on my fork and I've opened a draft PR (#46).

Please let me know if this is something you would be happy to merge if it was complete. Also, any feedback on the implementation would be greately appreciated