The empirical success of modern deep learning heavily relies on training on large datasets which may potentially contain private and sensitive data. During training, the parameters of a neural network are optimised based on samples from the training dataset. Previous works have shown that under certain conditions, reconstructing samples from the training data distribution is possible. This project proposes to extend previous works from shallow classification networks to problems in generative modelling. We also aim to gain a better understanding of the distribution of the reconstructions in comparison to the actual training distribution.
Training Data Reconstruction
My MS students work closely with stellar PhDs and PostDocs and world-class international collaborators. Some examples:
Røskva Bjørgfinsdóttir and Brage Eilertsen (first authors of AAAI submission over a summer project). first job after graduation: somewhere very decent!
Iman Tabrizian (co-first author of NeurIPS paper). first job after graduation: full-time engineer at NVIDIA
Wanyun Xie (first author of TMLR paper). first job after graduation: Ph.D. at EPFL
Published Nov. 2, 2025 7:38 AM
- Last modified Nov. 2, 2025 7:38 AM