626_privacy/tensorflow_privacy
2
0
Fork 1
Code Issues Pull requests 1 Projects Releases Packages Wiki Activity Actions

Prettier README

Browse source
This commit is contained in:
Nicholas Carlini 2021-12-14 00:54:29 +00:00
parent 7e40ad9704
commit a41549d103
1 changed files with 21 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

27
research/mi_lira_2021/README.md
View file

@ -1,3 +1,5 @@
## Membership Inference Attacks From First Principles
This directory contains code to reproduce our paper: This directory contains code to reproduce our paper:
**"Membership Inference Attacks From First Principles"** **"Membership Inference Attacks From First Principles"**
@ -5,7 +7,7 @@ https://arxiv.org/abs/2112.03570
by Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. by Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer.
###INSTALLING ### INSTALLING
You will need to install fairly standard dependencies You will need to install fairly standard dependencies
@ -17,9 +19,9 @@ https://github.com/google/objax
https://objax.readthedocs.io/en/latest/installation_setup.html https://objax.readthedocs.io/en/latest/installation_setup.html
###RUNNING THE CODE ### RUNNING THE CODE
####1. Train the models #### 1. Train the models
The first step in our attack is to train shadow models. As a baseline The first step in our attack is to train shadow models. As a baseline
that should give most of the gains in our attack, you should start by that should give most of the gains in our attack, you should start by
@ -45,7 +47,7 @@ exp/cifar10/
-- tb/ -- tb/
``` ```
####2. Perform inference #### 2. Perform inference
Once the models are trained, now it's necessary to perform inference and save Once the models are trained, now it's necessary to perform inference and save
the output features for each training example for each model in the dataset. the output features for each training example for each model in the dataset.
@ -65,7 +67,7 @@ where this new file has shape (50000, 10) and stores the model's
output features for each example. output features for each example.
####3. Compute membership inference scores #### 3. Compute membership inference scores
Finally we take the output features and generate our logit-scaled membership inference Finally we take the output features and generate our logit-scaled membership inference
scores for each example for each model. scores for each example for each model.
@ -84,7 +86,7 @@ exp/cifar10/
with shape (50000,) storing just our scores. with shape (50000,) storing just our scores.
###PLOTTING THE RESULTS ### PLOTTING THE RESULTS
Finally we can generate pretty pictures, and run the plotting code Finally we can generate pretty pictures, and run the plotting code
@ -112,3 +114,16 @@ where the global threshold attack is the baseline, and our online,
online-with-fixed-variance, offline, and offline-with-fixed-variance online-with-fixed-variance, offline, and offline-with-fixed-variance
attack variants are the four other curves. Note that because we only attack variants are the four other curves. Note that because we only
train a few models, the fixed variance variants perform best. train a few models, the fixed variance variants perform best.
### Citation
You can cite this paper with
```
@article{carlini2021membership,
title={Membership Inference Attacks From First Principles},
author={Carlini, Nicholas and Chien, Steve and Nasr, Milad and Song, Shuang and Terzis, Andreas and Tramer, Florian},
journal={arXiv preprint arXiv:2112.03570},
year={2021}
}
```