wiki:org/meetings/2018Rome/Notes/FloatingPointNoise

Random Floating Point

Generate uniformly random float with probability 1/2 in [0.5, 1.0), probability 1/4 in [0.25, 0.5), and so on…

Differential Privacy with Floating Point

Clamp, then add noise, then snap/bin, then clamp.

Randomising the low bits after scaling doesn't necessarily work: snapping correctly hides the low bits.

Floating Point Distributions

Error of sqrt(), cos(), and ln().

Do a floating-point error analysis on the alternative Gaussian transforms, and then pick one.

Use an argument similar to the one at the end of section 5.2 in: https://www.microsoft.com/en-us/research/wp-content/uploads/2012/10/lsbs.pdf

Use a numerical methods text for reference.

Last modified 4 months ago Last modified on Mar 15, 2018, 1:12:18 PM