human-analysis / maxent-arl

PyTorch Implementation of CVPR'19 (oral) - Mitigating Information Leakage in Image Representations: A Maximum Entropy Approach

MaxEnt-ARL: Mitigating Information Leakage in Image Representations: A Maximum Entropy Approach

By Proteek Chandan Roy and Vishnu Naresh Boddeti

Introduction

This code archive includes the Python implementation of MaxEnt-ARL for mitigating leakage of sensitive information from learned image representations.

Citation

If you think MaxEnt-ARL is useful to your research, please cite:

@article{roy2019mitigating,
    title={Mitigating Information Leakage in Image Representations: A Maximum Entropy Approach},
    author={Roy, Proteek Chandan and Boddeti, Vishnu Naresh},
    booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
    year={2019}
}

Usage

python main.py

By default this will run the experiment on CIFAR-100 dataset as described in the paper. Note that it will generate multiple runs of the trade-off between utility and privacy. The non-dominated solutions across the multiple runs provides the final trade-off front as reported in the paper.

In order to run the experiments on the other datasets in the paper, please edit the "main.py" file.

maxent-arl's Issues

Regarding cifar 100 results

Hello, I am running the main.py file which runs the cifar 100 experiments by default. The experiment runs for one time (outermost loop) for different values of alpha. In the end, it gives 71-72% accuracy for the target task. However, it gives around 20% accuracy for the privacy task which is different from the reported. It will be helpful if you can provide hyperparameters for reproducing the accuracy results on cifar 100 as mentioned in the supplementary material.