I am on a researcher position at Inria (Rennes), in the WIDE team.
“Maybe, not content to be alienated by observation, maybe does the object deceive us? Maybe it invented original answers, and not only the ones we ask for? May not be does it not want to be analyzed and observed at all and taking this for a challenge (which is true), does it respond with another challenge? […] the object analyzed today triumphs everywhere, by its object position, over the subject of analysis.” (Jean Baudrillard, Fatal Strategies.)
My main current research activity is trying to disprove that quote, by studying black-box algorithms in the context of recommender systems, neural-network models, or decision-making algorithms in general. I co-organized with Gilles and Antoine the workshop “Algorithmes en Boite-Noire”, in a continuous attempt to link several research domains on this societal question. e.g.:
News
- Launched the Elections 2022 sur YouTube website, to observe the presence of candidates in recommendations.
- YouTube Recommendations Do Predict Polls: A note on the 2022 French presidential election .
- Pre-print of the general approach to automate measurement, and to measure the “rabbit hole” effect of recommendations..
- Interview (in French) by Le Media TV on YouTube.
- I am co-chair of Eurosys workshops in Rennes in 2022.
- Article “Le problème du videur : la crédibilité des explications de l’IA en question” on interstices.info.
- I am the PI of the Audita associate team, btw EPFL (SaCS team) and Inria (WIDE), starting in 2021.
- I am at the scientific board of REGALIA.
- Paper accepted in CVPR 2021: SurFree: a fast surrogate-free black-box attack, presenting a fast (in number of queries) attack to find adversial examples.
- Paper accepted in INFOCOM 2021: Setting the Record Straighter on Shadow Banning, about the shadow banning practices in Twitter.
- Well mentioned in a The Atlantic paper on the topic: Shadowbanning Is Big Tech’s Big Problem.
- Paper cited as an evidence in a question to the European parliament “Censorship and free market restrictions including shadow banning and concealment or suppression of organic search results on internet”.
- press: Exploring the underpinnings of shadowbanning on Twitter TechXplore.
- Related to our black-box observations of YouTube recommendations
- New article in Nature Machine Intelligence, do not trust remote AI explanations… [Article link]
- pres: “Le problème du videur : la crédibilité des explications de l’IA en question” sur interstices.info.
- press: Do explanations for data-based predictions actually increase users’ trust in AI? Tech Xplore. Rebranded in Techhq.
- press: De l’explicabilité des systèmes : les enjeux de l’explication des décisions automatisées internetactu.net.
- press: Explainable AI” doesn’t work for online services – now there’s proof Algorithm Watch.
- press: It’s Too Easy to Hide Bias in Deep-Learning Systems IEEE spectrum.
- Juin 2020: participation à la task force “Régulation des GAFA / Transparence et audit des algorithmes” auprès de la Direction Générale des Entreprises (DGE).
- Semestre Supervision de systèmes dynamiques.
- Feb 2020: joined the board of the Société Informatique de France.
Recent publications (others in e.g., scholar)
SurFree: a fast surrogate-free black-box attack, Thibault Maho, Teddy Furon, Erwan Le Merrer. In CVPR (2021). Preprint [Arxiv 2020], [code].
RoBIC: A benchmark suite for assessing classifiers robustness, Thibault Maho, Benoît Bonnet, Teddy Furon, Erwan Le Merrer. In ICIP (2021) Preprint [Arxiv 2021], [code].
Setting the Record Straighter on Shadow Banning, Erwan Le Merrer, Benoît Morgan, Gilles Trédan. In INFOCOM (2021). [IEEE link] [ArXiv preprint 2020].
Remote Explainability faces the bouncer problem, Erwan Le Merrer, Gilles Trédan. In Nature Machine Intelligence (2020). Paper PDF doi [citation].
FeGAN: Scaling Distributed GANs, Rachid Guerraoui, Arsany Guirguis, Anne-Marie Kermarrec, Erwan Le Merrer. In Middleware, 2020. [Preprint], [Presentation video].
The Imitation Game: Algorithm Selection by Exploiting Black-Box Recommenders, Georgios Damaskinos, Rachid Guerraoui, Erwan Le Merrer and Christoph Neumann. In NETYS, 2020. [Preprint].
zoNNscan : a boundary-entropy index for zone inspection of neural models, Adel Jaouen, Erwan Le Merrer. In Monte Carlo Search workshop (MCS) @ IJCAI 2020 [ArXiv preprint 2018].
Adversarial Frontier Stitching for Remote Neural Network Watermarking, Erwan Le Merrer, Patrick Perez, Gilles Trédan. In Neural Comput & Applic 32, 9233–9244 (2020). [editor] [preprint] [citation]
Background
I was a senior research scientist at Technicolor R&I (2009-2018), where I worked on scalable storage, processing and machine learning for data analytics. I joined Inria in 2018, an on a permanent researcher position in 2021. I obtained my habilitation (HDR) on November 2016 from University of Rennes 1. I am the president and on of the sysadmins of the gozdata association, that provides the gozmail service, built from free software. I am a board member of the Société Informatique de France.
Profile photo © Inria / Photo B. Fourrier