Prof. Alexandre Graell i Amat (June 30, 2025 at 10:00 AM, Seminar room N2407, Zoom)
Talks |
Are You in the Training Set? Membership Inference and the Art of Attacking Machine Learning Models
Prof. Alexandre Graell i Amat
Chalmers University of Technology
Abstract:
Machine learning models are known to leak information about their training data. Because modern models are too complex for formal analysis, one of the most effective ways to understand their privacy risks is to actively build attacks against them.
Membership inference attacks (MIAs) are the most fundamental and widely studied of these: given a data point and atrained model, can an adversary tell whether the point was in the training set? MIAs serve as a kind of worst-case privacy test, where a strong adversary (one with near-complete knowledge of the data distribution) tries to infer just a single bit of information. Despite their simplicity, MIAs are explicitly mentioned in the European Data Protection Board guidance and underpin more sophisticated threats like data reconstruction.
In this seminar, we will look at how MIAs are built and what they reveal, with examples spanning both i.i.d. and graph-structured data. The talk will be accessible to all students (no machine learning background is assumed).
Biography:
Alexandre Graell i Amat is Professor with the Communication Systems group. His research interests are in the area of (modern) coding theory and cover a broad range of topics, including distributed storage, caching, and distributed computing, and optical communications.
Prof. Graell i Amat received the MSc degree in Telecommunications Engineering from the Universitat Politècnica de Catalunya, Barcelona, Catalonia, Spain, in 2001, and the MSc and the PhD degrees in Electrical Engineering from the Politecnico di Torino, Turin, Italy, in 2000 and 2004, respectively. Previously to joining Chalmers in 2011, he was with ENST Bretagne, Brest, France, from 2006 to 2010, first as a Marie Curie Fellow and then as an Associate Professor.
Prof. Graell i Amat is Editor at Large for the IEEE Transactions on Communications. Previously, he served as Associate Editor for the IEEE Transactions on Communications (2011-2016) and for the IEEE Communications Letters (2011-2013). He was the General Co-Chair of the 7th International Symposium on Turbo Codes & Iterative Information Processing, 2012, the Swedish Communication Technologies Workshop 2013, and the IEEE European School of Information Theory, 2016.
Prof. Graell i Amat is the recipient of the 2010 IEEE ComSoc Best Young Researcher Award for the Europe, Middle East, and Africa Region, and of a Marie Curie Intra-European Fellowship (2006-2008) by the European Commission. He is co-author of a paper that won the 2016 IEEE Sweden VT-COM-IT joint chapter best student conference paper award.