From db7711271299271784597d7260f63d27d2a18749 Mon Sep 17 00:00:00 2001 From: Jan Aalmoes Date: Fri, 13 Dec 2024 10:52:09 +0100 Subject: =?UTF-8?q?orhtographe=20derni=C3=A8re=20cerification?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- ckoi.tex | 65 ++++++++++++++++++++++++++++++++++++++++++++++------------------ 1 file changed, 47 insertions(+), 18 deletions(-) (limited to 'ckoi.tex') diff --git a/ckoi.tex b/ckoi.tex index 9922989..3dd2806 100644 --- a/ckoi.tex +++ b/ckoi.tex @@ -64,7 +64,7 @@ \begin{frame} \frametitle{L'IA comme produit} \begin{minipage}{270px} - \includegraphics[width=\linewidth]{images/ckoi/gafam.png} + \includegraphics[width=\linewidth]{images/ckoi/gafam2.png} \end{minipage} \begin{minipage}{120px} \footnotesize @@ -138,6 +138,11 @@ prédictions, du contenu, des recommandations ou des décisions qui peuvent infl \usebackgroundtemplate{\includegraphics[width=\paperwidth]{images/background/toc/background.pdf}} \begin{frame} +\Large +\underline{Introduction} +\pause +\normalsize + IA : enjeux, fonctionnement et modélisation de l'apprentissage automatique \vspace{10px} \pause @@ -146,6 +151,11 @@ prédictions, du contenu, des recommandations ou des décisions qui peuvent infl \vspace{10px} \pause +\Large +\underline{Contributions} +\pause +\normalsize + \emph{L'apprentissage ensembliste pour attaquer l'attribut sensible} \vspace{10px} \pause @@ -404,28 +414,39 @@ age : 85 sex : 1 race : 0 \begin{minipage}[t]{0.45\linewidth} \centering \input{tikz/attack_mia} - Inférence de l'appartenance aux données d'entraînement (MIA) + Inférence de l'appartenance aux données d'entraînement (MIA\textsuperscript{1}) \end{minipage} \hspace{10px} \begin{minipage}[t]{0.45\linewidth} \centering \input{tikz/attack} - Inférence d'un attribut sensible (AIA) + Inférence d'un attribut sensible (AIA\textsuperscript{2}) \end{minipage} + + \vspace{40px} + \footnotesize +1. \textit{Membership Inference Attack} + +2. \textit{Attribute Inference Attack} \end{frame} \begin{frame} \frametitle{Imposer l'équité} \emph{L'algorithme d'entraînement peut être modifié pour imposer l'équité au modèle final.} \begin{itemize} - \item \textit{FairGrad: Fairness Aware Gradient Descent}, Gaurav Maheshwari and Michaël Perrot, 2022. - \item \textit{Mitigating Unwanted Biases with Adversarial Learning}, Brian Hu Zhang and Blake Lemoine and Margaret Mitchell, 2018. - \item - \textit{A reductions approach to fair classification.} - Agarwal, A. and Beygelzimer, A. and Dudík, M. and Langford, J. and Wallach, H., 2018 - %\item \textit{Deep Learning with Differential Privacy}, - % Martín Abadi and Andy Chu and Ian Goodfellow, 2016. + \item Apprentissage adverse avec deux réseaux de neurones en compétition\textsuperscript{1} + \item Optimisation sous contrainte : approche problème primal/dual\textsuperscript{2,3} \end{itemize} + + \vspace{40px} + \footnotesize + 1. \textit{Mitigating Unwanted Biases with Adversarial Learning}, Brian Hu Zhang and Blake Lemoine and Margaret Mitchell, 2018. + + 2. \textit{FairGrad: Fairness Aware Gradient Descent}, Gaurav Maheshwari and Michaël Perrot, 2022. + + + 3. \textit{A reductions approach to fair classification.} + Agarwal, A. and Beygelzimer, A. and Dudík, M. and Langford, J. and Wallach, H., 2018 \end{frame} \begin{frame} @@ -433,16 +454,23 @@ age : 85 sex : 1 race : 0 \begin{itemize} \item \emph{Il y a un compromis à faire entre équité et MIA.} \begin{itemize} - \item \textit{On the Privacy Risks of Algorithmic Fairness}, - Hongyan Chang and Reza Shokri, 2021. - \item \textit{Differential Privacy Has Disparate Impact on Model Accuracy}, Eugene Bagdasaryan and Cornell Tech and Omid Poursaeed and Cornell Tech and Vitaly Shmatikov, 2019. + \item Imposer l'équité augmente le risque de MIA\textsuperscript{1} + \item Mitiger le risque de MIA avec de la confidentialité différentielle induit des inéquités \textsuperscript{2} \end{itemize} - \pause \item \emph{L'équité et l'AIA présentent des similarités.} \begin{itemize} - \item \textit{Mitigating Unwanted Biases with Adversarial Learning}, Brian Hu Zhang and Blake Lemoine and Margaret Mitchell, 2018. + \item Mitiger le risque AIA implique de rendre le modèle équitable\textsuperscript{3} \end{itemize} \end{itemize} + + \vspace{30px} + \footnotesize + 1. \textit{On the Privacy Risks of Algorithmic Fairness}, + Hongyan Chang and Reza Shokri, 2021. + + 2. \textit{Differential Privacy Has Disparate Impact on Model Accuracy}, Eugene Bagdasaryan and Cornell Tech and Omid Poursaeed and Cornell Tech and Vitaly Shmatikov, 2019. + + 3. \textit{Mitigating Unwanted Biases with Adversarial Learning}, Brian Hu Zhang and Blake Lemoine and Margaret Mitchell, 2018. \end{frame} \begin{frame} @@ -457,11 +485,12 @@ age : 85 sex : 1 race : 0 \caption{Après rééquilibrage adverse} \end{subfigure} \end{figure} - \textit{Mitigating Unwanted Biases with Adversarial Learning}, Brian Hu Zhang and Blake Lemoine and Margaret Mitchell, 2018. - \vspace{50px} + \vspace{40px} \footnotesize - \textsuperscript{1}\textit{Adversarial debiasing} + 1. \textit{Adversarial debiasing} + + \textit{Mitigating Unwanted Biases with Adversarial Learning}, Brian Hu Zhang and Blake Lemoine and Margaret Mitchell, 2018. \end{frame} \begin{frame} -- cgit v1.2.3