mahmoods01/accessorize-to-a-crime

Code for attacking state-of-the-art face-recognition system from our paper: M. Sharif, S. Bhagavatula, L. Bauer, M. Reiter. "Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition." In Proc. CCS, 2016.

Description

This repo contains the implementation of attacks on face-recognition systems from our CCS 2016 paper (see reference below). For the more recent work, please see: https://github.com/mahmoods01/agns.

Dependencies

• MatConvNet (http://www.vlfeat.org/matconvnet/), a convlutional neural networks library for Matlab.

• dlib for Python (https://pypi.python.org/pypi/dlib) is used for landmark localization and image alignment. Here's a good tutorial for how to install dlib: https://www.pyimagesearch.com/2017/03/27/how-to-install-dlib/.

How to run

To run the provided demo:

1. Run init.m to initialize the MatConvNet toolbox and add relevant paths.
2. Run demos of digital and physical attacks by running demo.m. The demos use the neural network that was provided by the VGG group (See: http://www.robots.ox.ac.uk/~vgg/software/vgg_face/).

The images under data/ have been aligned already, so one can run the demo before installing dlib. To align new images, the code in preprocess_data.m can be used.

Note: There are paths in init.m, demo.m, and preprocess_data.m that need to be updated (e.g., MatConvNet's installation path in init.m). The comment % update me was added to denote these paths.

Reference

If you use our code, please cite:

@inproceedings{Sharif16AdvML,
  author =       {Mahmood Sharif and Sruti Bhagavatula and Lujo Bauer 
						and Michael K. Reiter},
  title =        {Accessorize to a crime: {R}eal and stealthy attacks 
  						on state-of-the-art face recognition},
  booktitle =    {Proceedings of the 23rd ACM SIGSAC Conference on 
						Computer and Communications Security},
  year =         2016
}