This document supports the work The Many-faced God: Attacking Face Verification System with Embedding and Image Recovery (Accepted by ACSAC 2021 Annual Computer Security Applications). PAPER LINK
Author Invloved: Mingtian Tan,Zhe Zhou,Zhou Li
pyhton3 main.py -b 32 -l 0.0015
This file will conduct the whole Attack directly. "-b" means batch size and "-l" means learning rate. There are several variable settings you should pay attention, such as "datset path" or "model save path", which are related to your own project file structure.
Also, you can change the Hyper Parameters in the file independently or adjust the architecture of the model, which may result in better performance in face recovery task.
erGAN.py
This file is about the whole architecture of our erGAN model. Specifically, this file is about Embedding-1024 Face Modelrecovery in real world, also you can change the model interface to fit your own recover task. "_generator()" is about how we extract information from embedding and recover face from it, pipline showing below:
This is the performance ranmdomly choosed from Testing Data. The first line is the oringial face images from public dataset "LFW". Second line is the face images recovered from embedding 1024 from an online face classification application Clarifai-1024.
Do NOT use the contents of this repository in applications which handle sensitive data. The author accepts no liability for privacy infringements - use the contents of this repository solely at your own discretion. We conduct all the experiments in public data set, such as LFW or CASIA.
We will continue to update this project later. If you are intesested in our project please contact us here or send email to me (mttan@smu.edu.sg). We welcome your communication very much.