search

bargavj / EvaluatingDPML

This project's goal is to evaluate the privacy leakage of differentially private machine learning models.
MIT License
129 stars 48 forks source link

Fix reordering of pred_scores aka attack_pred #12

Closed jonahweissman closed 5 years ago

jonahweissman commented 5 years ago

Because a separate attack model is trained for each output class, the predictions for the attack testing set were grouped by output class, instead of the usual format. I used the indices to place the predictions where they belong.

jonahweissman commented 5 years ago

Also, I should note that attack_pred now just represents the prediction of a datapoint being a member, instead of also including the complement.

i.e. new_attack_pred = old_attack_pred[: , 1]

bargavj commented 5 years ago

I see. That makes sense. I will go ahead and merge the pull request.