Open adam-dziedzic opened 3 years ago
Thanks for spotting this issue! I have updated the corresponding code in the new commit. I tried to rerun the experiments on CelebA to compare the effects induced by tau (clipping threshold). The accuracy is now dropped to 84.70% (slightly worse than what I reported in the paper), when tau =10, k=800, sigma = 50, number of queries = 800 and sampling ratio = 0.05. If tau is set to be 5 and other hyper-parameters remain the same, the accuracy is around 84.0% while privacy cost epsilon <1.0. The feature is based on ImageNet pre-trained Resnet50m (using hog-based feature will drop the accuracy by around one percent). Hope this would be useful for you.
Thanks for fixing it. I added a comment to the code: Shouldn't the teachers_preds
be replaced with preds_tau
as well for the count_zero_list?
Hello,
Thanks for sharing the code. It seems that despite setting config.use_tau to True, the full votes
teachers_preds
are used instead of the clipped votespreds_tau
:https://github.com/jeremy43/Private_kNN/blob/f321d0c32a4bd19a56fa488152403e0c62379bc3/face_attribute/knn_attribute.py#L174