Skip to content

Latest commit

 

History

History
32 lines (23 loc) · 2.06 KB

README.md

File metadata and controls

32 lines (23 loc) · 2.06 KB
Raw

Adversarial Patch Attacks on Face Recognition Neural Networks

As part of the 2018 NSF-Funded Computer Science REU at the University of Maryland, me (Ryan Synk) and a team of undergraduates studied adversarial attacks on facial recognition neural networks. In September 2020, I heavily updated the repository — originally located on the University of Maryland Institute for Advanced Computer Studies (UMIACS) Gitlab — and migrated it onto my personal github page.

This repository contains code written by Ryan Synk, Lara Shonkwiler, Kate Morrison, Xinyi Wang and Carlos Castillo. Credit for this work is spread evenly among those names listed, as well as Prof. Tom Goldstein, who directed the group.

Directions to Generate Adversarial Patch

The generate_patch.py script takes a neural network, a base image, and a target image, and optimizes a small, square patch. This patch, when overlayed on top of the base image, causes the network to interpret the image as the target. To generate a patch, simply clone the repository and run the "generate_patch.py" script. To run the script with images of your choosing, pass them in as parameters on the command line. The default should show something like this

Directions for Generating Images

This software uses UMDFaces and Pytorch to train deep networks for face recognition.

The steps are:

  1. Generate thumbnails to train. Use compute_aligned_images.py for this task, point to the three batches of UMDFaces, run it three times.
  2. Copy the create.py script to the val directory and run it there. This will create the missing directories in val that are required for validation to work.
  3. Train. I used python main.py --pretrained --epochs 200 --lr 0.1 --print-freq 1 /scratch2/umdfaces-thumbnails/. You should come back 12 hours later.
  4. Generate features, for example. I used the compute_features.py script for this task.
  5. Generate plots/statistics, whatever you want, I used run_lfw.py for this.