Improving Robustness to Model Inversion Attacks via Mutual Information Regularization

This work is appeared in AAAI'21. Paper preprint is available at https://arxiv.org/abs/2009.05241.

Requirements (not too important as long as you can run)

torch==1.0

python train_inv.py used to train MI attack models. python train_vib.py and python train_dp are used to train MID and DP models, respectively.

For Facescrub dataset, we provide a very small subset for it. To downlaod the full dataset, please refer to http://vintage.winklerbros.net/facescrub.html.

Name		Name	Last commit message	Last commit date
Latest commit History 4 Commits
gmi		gmi
knowledge_align		knowledge_align
update_leaks		update_leaks
.DS_Store		.DS_Store
README.md		README.md