Hi,
I have a question regarding how the score for the offline attack is calculated in the Research mi_lira_2021 code.
In the code the score is calculated using logpdf
score = scipy.stats.norm.logpdf(sc, mean_out, std_out+1e-30)
but if I understand correctly this does not match equation 4 in the paper that describes how the score should be calculated. The equation looks like this $\Lambda = 1- Pr[Z > \phi (f(x)_{y} ) ]$, where $Z \sim \mathcal{N} (\mu_o, \sigma^2_o)$ which should mean that cdf should be used and not pdf. Is there any particular reason for why the pdfis used in the implementation over cdf?
Hi,
I have a question regarding how the score for the offline attack is calculated in the Research mi_lira_2021 code.
In the code the score is calculated using logpdf
but if I understand correctly this does not match equation 4 in the paper that describes how the score should be calculated. The equation looks like this$\Lambda = 1- Pr[Z > \phi (f(x)_{y} ) ]$ , where $Z \sim \mathcal{N} (\mu_o, \sigma^2_o)$ which should mean that
cdfshould be used and notpdf. Is there any particular reason for why thepdfis used in the implementation overcdf?