ACSAC/figures/fig_egd_attack.tex


1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50


\begin{figure}[!htb]
    \centering
    \begin{minipage}[b]{1\linewidth}
    \centering
    \subfigure[\census (\race)]{
    \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/census/census_egd_attack_hard_race.pdf}
    }%
    \subfigure[\census (\sex)]{
    \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/census/census_egd_attack_hard_sex.pdf}
    }
    \end{minipage}\\

   \begin{minipage}[b]{1\linewidth}
   \centering
   \subfigure[\compas (\race)]{
   \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/compas/compas_egd_attack_hard_race.pdf}
   }%
   \subfigure[\compas (\sex)]{
   \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/compas/compas_egd_attack_hard_sex.pdf}
   }
   \end{minipage}\\

    \begin{minipage}[b]{1\linewidth}
    \centering
    \subfigure[\meps (\race)]{
    \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/meps/meps_egd_attack_hard_race.pdf}
    }%
    \subfigure[\meps (\sex)]{
    \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/meps/meps_egd_attack_hard_sex.pdf}
    }
    \end{minipage}\\

   \begin{minipage}[b]{1\linewidth}
   \centering
   \subfigure[\lfw (\race)]{
   \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/lfw/lfw_egd_attack_hard_race.pdf}
   }%
   \subfigure[\lfw (\sex)]{
   \includegraphics[width=0.49\linewidth]{ACSAC/figures/egd/lfw/lfw_egd_attack_hard_sex.pdf}
   }
   \end{minipage}

    \caption{For \adaptiveAIAHard, we observe that \egd reduces the attack accuracy to random guess ($\sim$50\%). %Additionally, the theoretical bound on attack accuracy (``Theory'') matches with the empirical results (``Empirical'').
    }
    \label{fig:AdaptAIAEGD}
\end{figure}


%\textbf{Overall, we observe that using group fairness results in attribute privacy but comes at the cost of $\targetmodel$'s utility.}