Learning from Data (Fall 2022)

[kundtx]

http://8.129.175.102/lfd2022fall-poster-session/42.html

[NYZTTTT]

G4 Zihao Ni: By how much did you reduce the expence of privacy budget? Is there any metric used to evaluate the new algorithm on the privacy and utility?

[Caarinaaa]

G10 Jinnan He: What are the PARAMETERS listed in the RESULTS part? And what do they imply? I am a bit confused. Thx!

[tsld21]

@NYZT1027 G4 Zihao Ni: By how much did you reduce the expence of privacy budget? Is there any metric used to evaluate the new algorithm on the privacy and utility?

Actually the measure of budget and the metric of utility and privacy are still unsettled problems, I mean there is no standard principle. \ 1)So for your first question, here I just finished that in my experiment, which is the $\epsilon$ used to adjust privacy degree in differential privacy. Honestly, here we need some experiments, as well as mathematical proof to prove this. \ 2)For your second question, as I said, we don't have an acknowledged metrics to measure the balance. Usually, for privacy, we can use some attack model to evaluate. And for utility, from the perspective of ML, we can use MSE or some statistical metrics, such as mean, variance, also we can adopt KL divergence to evaluate the utility.\ And this poster illustrate a new approach combining exponential mechanism with ML, so I am still working on it.

[tsld21]

@Caarinaaa G10 Jinnan He: What are the PARAMETERS listed in the RESULTS part? And what do they imply? I am a bit confused. Thx!

Thanks for your question, here I mean machine learning model parameters, theta. Since this is a totally new method, I have been working on it to show that's feasible. So here I used data composed of 100 records with 2 variables, so the parameters here is a 2-dimension array. \ Sorry for making you confused.