Confidence is All You Need for MI Attacks

This directory contains code to reproduce our paper: "Confidence is all you need for MI Attacks"
https://arxiv.org/abs/2311.15373
by Abhishek Sinha, Himanshi Tibrewal, Mansi Gupta, Nikhar Waghela, Shivank Garg

Our work is based upon :

"Membership Inference Attacks From First Principles"
https://arxiv.org/abs/2112.03570
by Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramèr.

INSTALLING DEPENDENCIES

To install the basic dependencies needed to run this repository

bash requirements.sh

We train our models with JAX + ObJAX so you will need to follow build instructions for that https://github.com/google/objax https://objax.readthedocs.io/en/latest/installation_setup.html

RUNNING THE CODE

1. Train the models

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 training 16 shadow models with the command

bash scripts/train_demo.sh

or if you have multiple GPUs on your machine and want to train these models in parallel, then modify and run

bash scripts/train_demo_multigpu.sh

This will train several CIFAR-10 wide ResNet models to ~91% accuracy each, and will output a bunch of files under the directory exp/cifar10 with structure:

exp/cifar10/
- experiment_N_of_16
-- hparams.json
-- keep.npy
-- ckpt/
--- 0000000100.npz
-- tb/

2. Perform inference

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.

python3 inference.py --logdir=exp/cifar10/

This will add to the experiment directory a new set of files

exp/cifar10/
- experiment_N_of_16
-- logits/
--- 0000000100.npy

where this new file has shape (50000, 10) and stores the model's output features for each example.

3. Compute membership inference scores

Finally we take the output features and generate our logit-scaled membership inference scores for each example for each model.

python3 score.py exp/cifar10/

We find the evaluation of scores through various experiments. The calculations of logits are implemented in the score.py file, where we explored all the commented-out calculations to find the logits. It was noted that utilizing argmax values, which doesn't require knowledge of true labels, produced results comparable to those outlined in the "LIRA Likelihood Ratio Paper."

And this in turn generates a new directory

exp/cifar10/
- experiment_N_of_16
-- scores/
--- 0000000100.npy

with shape (50000,) storing just our scores.

PLOTTING THE RESULTS

Finally we can generate pretty pictures, and run the plotting code.

python3 plot.py

The results will look similar to this

RESULTS {Using AUC as Metric}

	Loss Value (Baseline)	Confidence Values	log (Confidence Values)	Argmax	log (Argmax)
Attack Ours (Online)	0.5753	0.5668	0.575	0.5464	0.5447
Attack Ours (Online,Fixed Variance)	0.5879	0.593	0.6009	0.5622	0.5602
Attack Ours (Offline)	0.5181	0.492	0.4721	0.478	0.4756
Attack Ours (Offline, Fixed Variance)	0.5184	0.4928	0.4804	0.4834	0.4815
Attack Global Threshold	0.5448	0.5439	0.5469	0..5376	0.5377

Where the global threshold attack is the baseline, and our online, online-with-fixed-variance, offline, and offline-with-fixed-variance attack variants are the four other curves. Note that because we only train a few shadow models, the fixed variance variants perform best.

Citation

You can cite this paper with

@ title= {Confidence is All You Need For MI Attacks}
  author={Abhishek Sinha, Himanshi Tibrewal, Mansi Gupta, Nikhar Waghela, Shivank Garg},
  journal={arXiv preprint arXiv:2311.15373},
  year={2023}
}