/dataset_reconstruction

Primary LanguageJupyter Notebook

Reconstructing Training Data
from Trained Neural Networks

Niv Haim*, Gal Vardi*, Gilad Yehudai*, Ohad Shamir, Michal Irani

NeurIPS 2022 (Oral)

Webpage, Paper

Pytorch implementation of the NeurIPS 2022 paper: Reconstructing Training Data from Trained Neural Networks.

Setup

Create a copy of setting.default.py with the name setting.py, and make sure to change the paths inside to match your system.

Create and initialize new conda environment using the supplied environment.yml file (using python 3.8 and pytorch 11.1.0 and CUDA 11.3) :

conda env create -f environment.yaml
conda activate rec

Running the Code

Notebooks

For quick access, start by running the provided notebooks for analysing the (already provided) reconstructions of two (provided) models for binary CIFAR10 (vehicles/animals) and MNIST (odd/even):

  • reconstruction_cifar10.ipynb
  • reconstruction_mnist.ipynb

Reproducing the provided trained models and their reconstructions

All training/reconstructions are done by running Main.py with the right parameters.
Inside command_line_args directory we provide command-lines with necessary arguments for reproducing the training of the provided models and their provided reconstructions (those that are analyzed in the notebooks)

Training

For reproducing the training of the provided two trained MLP models (with architecture D-1000-1000-1):

  • CIFAR10 model (for reproduction run command_line_args/train_cifar10_vehicles_animals.txt)
  • MNIST model (for reproduction run command_line_args/train_mnist_odd_even_args.txt)

Reconstructions

In reconstructions directory we provide two reconstructions (results of two runs) per each of the models above.

To find the right hyperparameters for reconstructing samples from the above models (or any other models in our paper) we used Weights & Biases sweeps. In general, it is still an open question how to find the right hyperparameters for our losses without trial and error.

These reconstructions can be reproduced by running the following commandlines (the right hyperparameter can be found there):

  • CIFAR10: command_line_args/reconstruct_cifar10_b9dfyspx_args.txt and command_line_args/reconstruct_cifar10_k60fvjdy_args.txt
  • MNIST: command_line_args/reconstruct_mnist_kcf9bhbi_args.txt and command_line_args/reconstruct_mnist_rbijxft7_args.txt

Training/Reconstructing New Learning Problems

One should be able to train/reconstruct models for new problems by adding a new python file under problems directory.

Each problem file should contain the logic of how to load the data and the parameters necessary to build the model.

BibTeX

@article{haim2022reconstructing,
  title={Reconstructing Training Data from Trained Neural Networks},
  author={Haim, Niv and Vardi, Gal and Yehudai, Gilad and Shamir, Ohad and Irani, Michal},
  year={2022},
  eprint={2206.07758},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}