summaryrefslogtreecommitdiff
path: root/aia.tex
blob: 76cb80e5cfb015b4e80660a14aac85ef6aa2357a (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
107
108
109
110
111
112
\begin{frame}
    \begin{definition}[DemParLvl]
        Soit $(\Omega,\mathcal{T},P$) un espace probabilisé.
        Soit $(E,\mathcal{E})$ un espace mesurable.
        Soient
        \begin{align*}
            X:(\Omega,\mathcal{T})\rightarrow(E,\mathcal{E})\\
            Y:(\Omega,\mathcal{T})\rightarrow(\{0,1\},\mathcal{P}(\{0,1\}))\\
            S:(\Omega,\mathcal{T})\rightarrow(\{0,1\},\mathcal{P}(\{0,1\}))\\
        \end{align*}
        Soit $f:\{0,1\}\rightarrow\{0,1\}$.
        Alors,
        \begin{equation*}
            DemParLvl(f) = |P(f\circ X=0\mid S=0) - P(f\circ X=0\mid S=1)|
        \end{equation*}
    \end{definition}
    \pause
    \begin{propriete}
        \label{prop:aia-dpl0}
        Un classifieur qui satisfait la parité démographique a un DemParLvl égal à zéro.
    \end{propriete}
\end{frame}
\begin{frame}
    \begin{definition}[Parité démographique généralisée]
        \label{def:aia-dempargen}
    Soit $(\Omega,\mathcal{T},P$) un espace probabilisé.
    Soient $(E,\mathcal{E})$, $(F,\mathcal{F})$ et $(G,\mathcal{G})$ des espaces mesurables.
    Soient les variables aléatoires suivantes : 
    \begin{itemize}
        \item $X:(\Omega,\mathcal{T})\rightarrow (E,\mathcal{E})$
        \item $Y:(\Omega,\mathcal{T})\rightarrow (F,\mathcal{F})$
        \item $S:(\Omega,\mathcal{T})\rightarrow (G,\mathcal{G})$
        \item $f:(E,\mathcal{E})\rightarrow (F,\mathcal{F})$
    \end{itemize}
    Alors $f$ satisfait la parité démographique généralisée si et seulement si 
    \begin{equation*}
        P_{f\circ X,S} = P_{f\circ X}\otimes P_S
    \end{equation*}
    Dit autrement, si et seulement si le classifieur $f$ est un CCA pour prédire $S$ à partir de $X$.
    \end{definition}
\end{frame}

\begin{frame}
    \begin{propriete}
        Si un classifieur binaire satisfait la parité démographique généralisée alors il satisfait la parité démographique.
    \end{propriete}
\end{frame}

\begin{frame}

    \begin{theorem}
        \label{th:aia-dpgood}
        Les deux propositions suivantes sont équivalentes :
        \begin{enumerate}
            \item Le modèle cible satisfait la parité démographique .
            \item Toutes les attaques utilisant la prédiction pour inférer l'attribut sensible sont des CCA.
        \end{enumerate}

        Et aussi, les deux propositions suivantes sont équivalentes :
        \begin{enumerate}
            \item Le modèle cible satisfait la parité démographique généralisée.
            \item Toutes les attaques utilisant le logit pour inférer l'attribut sensible sont des CCA.
        \end{enumerate}
    \end{theorem}
\end{frame}

\begin{frame}
    \begin{propriete}
        \label{prop:aia-demparlvl}
        Soient $(\Omega,\mathcal{T},P)$ un espace probabilisé et $(\{0,1\}$, $\mathcal{P}(\{0,1\}))$ des espaces mesurables.
        Soient les variables aléatoires suivantes
        \begin{itemize}
            \item L'étiquette $Y:(\Omega,\mathcal{T})\rightarrow (\{0,1\},\mathcal{P}(\{0,1\}))$
            \item La donnée d'entrée $X:(\Omega,\mathcal{T})\rightarrow(\{0,1\},\mathcal{P}(\{0,1\})$
            \item L'attribut sensible $S:(\Omega,\mathcal{T})\rightarrow(\{0,1\},\mathcal{P}(\{0,1\}))$
            \item L'attaque $a:(\Omega,\mathcal{T})\rightarrow(\{0,1\},\mathcal{P}(\{0,1\}))$
            \item Le modèle cible $f:(\Omega,\mathcal{T})\rightarrow(\{0,1\},\mathcal{P}(\{0,1\}))$
        \end{itemize}
        Alors nous avons 
        \begin{equation*}
            \text{max}_{a}BA(a) = \frac{1}{2}(1+\text(DemParLvl(f)))
        \end{equation*}
    \end{propriete}
\end{frame}

\begin{frame}
    \begin{theorem}
        \label{th:aia-bluey}
        Soit $(\Omega,\mathcal{T},P)$ un espace probabilisé.
        Soient $(E,\mathcal{E})$ et $(F,\mathcal{P}(F))$ des espaces mesurables avec $F$ un ensemble fini.
        Soient les variables aléatoires suivantes :
        \begin{itemize}
            \item $X:\Omega\rightarrow E$
            \item $Y:\Omega\rightarrow F$
        \end{itemize}
        Soit $A$ l'ensemble des fonctions mesurables de $(E,\mathcal{E})$ dans $(F,\mathcal{P}(F))$.
        Nous appelons $BA$ la fonction qui à toutes les fonctions $a$ de $A$ associe l'exactitude équilibrée de $a \circ X$ pour l'étiquette $Y$.
        \begin{equation*}
        \exists a\in A~BA(a)< \frac{1}{\#F}
        \implies
        \exists a\in A~BA(a)>\frac{1}{\#F}
        \end{equation*}
    \end{theorem}
\end{frame}

\begin{frame}
    \begin{figure}
        \centering
        \input{tikz/data}
        \label{fig:aia-data}
    \end{figure}
\end{frame}