Experiments on post-facto methods inspired by Differential Privacy to protect BERT embeddings from inversion attacks while keeping the utility intact. The project explores the tradeoff between privacy and utility .
research pytorch noise bert differential-privacy mathematical-methods finetuning embedding-inversion-attack dp-forward
-
Updated
Aug 19, 2025