Adversarial Eyeglasses to Trick Facial Recognition

I recently read a paper by Sharif et al. that describes a general framework for adversarial example generation and they utilize eyeglass frames affixed to people’s faces to trick a facial recognition classifier. In addition, this method also worked when printing out the eyeglass frames and using them in a physical adversarial attack.

I decided to build a GitHub repository to implement this method using PyTorch. I also added some additional bells and whistles for inference (like trying the glasses out automatically on your webcam).