Achraf

[#PhDLife] Peut-on entraîner efficacement des algorithmes de prise de décision tout en assurant la confidentialité des données utilisées? La #thèse d'@achraf_azize (@InriaScool) s'attaque à ce problème dans le cas des essais cliniques. 👏Bravo à lui pour sa soutenance!

1/2 To define privacy in bandits, we have to ask what are the input and output of a bandit algorithm? What differs if the adversary is interactive or passive? @achraf_azize & @BasuDebabrota address these in their work openreview.net/forum?id=2366a….

2/2 Today @satml_conf, @achraf_azize is presenting these nuances of "Concentrated DP for Bandits" along with information-theoretic lower bounds and near-optimal algorithms for linear, contextual, and multi-armed bandits. #privacy #bandits @Inria_Lille @RechercheUlille #SaTML24

1/2 What's the cost of privacy in bandits? We prove that it's almost free 😉 At 10:45AM @NeurIPSConf, @achraf_azize, @MarcJourdan5 & @BasuDebabrota are going to present the first lower bounds and an efficient algorithm for private best arm identification. arxiv.org/abs/2309.02202

@NeurIPSConf, @InriaScool members are going to present 6 papers on #RL, #bandits, and #privacy 🎓We invite the attendees of #NeurIPS2023 to visit & enrich us with your questions! Also, contact Scoolmates at NOLA if you want to know more about works & opportunities in Scool🙏

On the Complexity of Differentially Private Best-Arm Identification with Fixed Confidence w/ A. Azize, @aymen_marjani & @BasuDebabrota Paper: arxiv.org/abs/2309.02202 Poster: Wed 13 Dec 10:45 a.m - 12:45 pm CST (#1604) 3/4

@TedOnPrivacy That's what I used in the section 3.2 of my answer to give an easy upper bound on this kl, and thus a (vi, xi) approx group wise zCDP where both vi and xi are both non increasing functions of x(D)_i

@TedOnPrivacy Hello Ted, there exists a closed form of Rényi divergence between two gaussians of different variance, see for example page 6 of mast.queensu.ca/~communication….

And now for something different: during a customer engagement, we stumbled upon a natural, but surprisingly tricky open problem in differential privacy. We couldn't figure it out, so we posted a simplified version of it as an open problem on DifferentialPrivacy.org 👀

@TedOnPrivacy I don't understand why this algorithm is not directly a solution to your problem.

@TedOnPrivacy I also analysed just the naive algorithm that adds gaussian noise with mean x_i and variance r^2 x_i^2, which can be shown to have the exact utility wanted + a privacy budget non-increasing in x_i (specifically in 1/x_i^2).