Introduction

The InstaHide framework aims to provide privacy-preserving machine learning by encoding sensitive datasets. However, our research reveals significant vulnerabilities in this approach. In this blog post, we will explore the implementation of a reconstruction attack on InstaHide, demonstrating how it fails to offer meaningful privacy protection.

Project Purpose and Main Features

The primary goal of this project is to showcase the weaknesses in the InstaHide framework. The main features of our implementation include:

• Reconstruction Attack: Recover original images from encoded datasets.
• Sequential Steps: A series of scripts to execute the attack effectively.
• Visualization: Tools to visualize the reconstructed images.

Technical Architecture and Implementation

The attack consists of several steps that must be executed in order. Each step builds upon the previous one, utilizing pre-trained neural network classifiers. Below is a brief overview of the implementation:

1. step_1_create_graph.py: Generates a similarity graph for encoded images.
2. step_2_color_graph.py: Colors the graph to identify dense cliques.
3. step_3_second_graph.py: Constructs a new bipartite similarity graph.
4. step_4_final_graph.py: Solves the matching problem for image assignment.
5. step_5_reconstruct.py: Reconstructs the original images.
6. step_6_adjust_color.py: Adjusts color curves for accuracy.
7. step_7_visualize.py: Displays the final reconstructed images.

Setup and Installation Process

To run the reconstruction attack, follow these steps:

1. Ensure you have Python 3.5+ installed.
2. Install the required libraries:

pip install jax jaxlib objax pillow scikit-learn

3. Download the necessary dependency files:
• encryption.npy
• labels.npy
• saved models
4. Run the scripts in the order specified above.

Usage Examples and API Overview

After setting up the environment and downloading the necessary files, you can execute the scripts to perform the reconstruction attack. Here’s a brief example of how to run the first step:

python step_1_create_graph.py

This command will generate the similarity graph needed for the subsequent steps. Each script is designed to be run sequentially, ensuring that the output of one step serves as the input for the next.

Community and Contribution Aspects

We welcome contributions to this project, especially in the research folder. If you have novel attacks or improvements, please follow our contribution guidelines. Ensure your code adheres to the Google Python style guide and includes proper documentation.

License and Legal Considerations

This project is licensed under the MIT License, allowing for free use, modification, and distribution. However, please ensure that you comply with the license terms when using or contributing to this project.

Conclusion

The InstaHide framework presents a significant challenge in the realm of privacy-preserving machine learning. Our reconstruction attack demonstrates that it does not provide adequate protection for sensitive data. By following the steps outlined in this blog post, you can replicate our findings and contribute to the ongoing discussion about privacy in machine learning.

Resources

For more information, check out the following resources:

• Research Paper
• GitHub Repository

FAQ