Protection of Deep Neural Networks for Face Recognition with Adversarial Attacks

Artificial Intelligence (AI) is a transformative technology with the potential to revolutionize our entire lives. The discovery of deep neural networks (DNNs), which give computers the ability to outperform humans in many things, including face recognition, was a milestone that has made AI so inviting. However, scientists have discovered very recently that DNN networks are especially vulnerable to attacks, i.e., the so-called adversarial examples, which are imperceptible to humans but can fool DNNs easily. In this project, a novel approach is proposed to defend against these adversaries, which is much more efficient than the often-used defense methods based on adversarial learning. The newly proposed idea utilizes the attacks themselves as a defense mechanism. The task of face recognition is used to experimentally validate the novel idea and approaches. Based on a large dataset with 6,000 pairs of face images, this new defense handles the adversarial attacks efficiently, improving the face recognition accuracies from about 0% under attack to over 80% after defense.