Defence for Face-morphing Adversarial Attacks on Facial Recognition Systems

  • Achieved a 200x better MSE loss on unknown faces, by utilizing a discriminator trained on morphed face images
  • Obtained 80% success rate, on attacking SOTA Face Recognition models, OpenFace and FaceNet512, by morphing two face images by interpolating their semantic and stochastic embeddings produced by Diffusion Autoencoders

Download Slides ; Download Report ; Github Repository