avnacé aia, remerciement notations, notes

1 files changed, 50 insertions, 0 deletions

diff --git a/ACSAC/figures/fig_egd_attack.tex b/ACSAC/figures/fig_egd_attack.tex
new file mode 100644
index 0000000..8b1c713
--- /dev/null
+++ b/ACSAC/figures/fig_egd_attack.tex
@@ -0,0 +1,50 @@
+
+\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.}
\ No newline at end of file