AGNs

Description

This repo contains an implementation of the Adversarial Generative Nets (AGNs), which we proposed in our ACM TOPS 2019 paper (see reference below). A demo for launching impersonation and dodging attacks against the VGG and OpenFace face recognition neural networks is provided.

Data and models

Before you can run the code, you need to download the set of eyeglass images that we synthesized from real textures (link: https://tinyurl.com/AGNsEyeglasses; place them under data/eyeglasses) and the neural networks that we trained (link: https://tinyurl.com/AGNsModels; place them under models/).

Dependencies

The code is implemented in MATLAB (we used MATLAB R2015a). As mentioned in the paper, our implementation depends on MatConvNet (http://www.vlfeat.org/matconvnet/)---a MATLAB toolbox for convolution neural networks. An extended version (containing additional layers, etc.) is provided under dependencies/.

To align images (necessary when running experiments with the OpenFace neural networks and for using new images in attacks), our code depends on Python packages (we used python3.6) for face and landmark detection. Specifically, the face and landmark detectors of the dlib package are used.

Instructions for running

The attack code (see under code/agn*.m) takes as input face images that are aligned to VGG's canonical pose. For physical attacks, the aligned images should also contain green marks that are used for aligning the eyeglasses to the face image (see examples in data/demo-data2/). See demo.m for examples of how to run attacks.

For the purpose of the demo, aligned images are provided with the code. To align new images, you can use the face- and landmark- detection, and face-alignment code under dependencies/image-registration/. Before running the code, you need to update the paths in the files face_landmark_detection.m and openface_align.m. See align_demo.m for a face-alignment example.

Reference

If you use the code, please cite our paper:

@article{Sharif19AGNs,
  author =       {Mahmood Sharif and Sruti Bhagavatula and 
  					  Lujo Bauer and Michael K. Reiter},
  title =        {A general framework for adversarial examples 
  					  with objectives},
  journal =      {ACM Transactions on Privacy and Security},
  year =         2019
}