blob: 665174492449d4f1597a1e5b074b2541854d408f (
plain)
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
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
|
{
\usebackgroundtemplate{\includegraphics[width=\paperwidth]{images/background/card/background.pdf}}
\begin{frame}
%\vspace{70px}
\hspace{70px}
\begin{minipage}{250px}
\Large
\textcolor{accent}{
Evaluation expérimentale de l'utilisation de l'équitée comme mécanisme de protéction de l'attribut sensible.
}
\end{minipage}
\end{frame}
}
\begin{frame}
\begin{figure}
\centering
\input{tikz/data}
\label{fig:aia-data}
\end{figure}
\end{frame}
\begin{frame}
\frametitle{Experimental validation on prediction: results}
\begin{figure}
\captionsetup{singlelinecheck=off}
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\scriptsize
\begin{itemize}
\item \emph{Labeled Faces in the Wild (images)}
\item ML = Convolutional Neural Network
\end{itemize}
\includegraphics[width=150px]{images/figures/advdebias/lfw/lfw_advdeb_attack_hard_sex.pdf}
\end{subfigure}
\hspace{0.1\textwidth}
\begin{subfigure}{0.4\textwidth}
\centering
\scriptsize
\begin{itemize}
\item \emph{COMPAS recidivism dataset (tabular)}
\item ML = Random Forest
\end{itemize}
\includegraphics[width=150px]{images/figures/advdebias/compas/compas_advdeb_attack_hard_sex.pdf}
\end{subfigure}
\end{figure}
\vspace{10px}
\scriptsize
\begin{tabular}{lll}
&\emph{Regularization}&\emph{Value}\\
\emph{Baseline}&None&Attack result\\
\emph{Theoretical}&Adversarial debiasing&$\frac{1}{2}(1+DemParLvl)$\\
\emph{Empirical}&Adversarial debiasing&Attack result\\
\end{tabular}
\normalsize
\hspace{10px}
Attack surface = $1_{[\tau,1]}\circ f\circ X$.
\end{frame}
\begin{frame}
\frametitle{Experimental validation on logit: building an attack}
\begin{enumerate}
\item On part.
\item Build a random forest on this dataset.
\item Ajust the threshold to take into account class imbalance.
\end{enumerate}
\end{frame}
\begin{frame}
\frametitle{Experimental validation on logit: results}
\begin{figure}
\captionsetup{singlelinecheck=off}
\centering
\begin{subfigure}{0.4\textwidth}
\centering
\scriptsize
\begin{itemize}
\item \emph{Labeled Faces in the Wild (images)}
\item ML = Convolutional Neural Network
\end{itemize}
\includegraphics[width=150px]{images/figures/advdebias/lfw/lfw_advdeb_attack_soft_experimental_sex.pdf}
\end{subfigure}
\hspace{0.1\textwidth}
\begin{subfigure}{0.4\textwidth}
\centering
\scriptsize
\begin{itemize}
\item \emph{COMPAS recidivism dataset (tabular)}
\item ML = Random Forest
\end{itemize}
\includegraphics[width=150px]{images/figures/advdebias/compas/compas_advdeb_attack_soft_experimental_sex.pdf}
\end{subfigure}
\end{figure}
\vspace{10px}
\scriptsize
\begin{tabular}{lll}
&\emph{Regularization}&\emph{Value}\\
\emph{Baseline}&None&Attack result\\
\emph{AdvDebias}&Adversarial debiasing&Attack result\\
\end{tabular}
\normalsize
\hspace{10px}
Attack surface = $f\circ X$.
\end{frame}
|
