Adversarial Training for Face Recognition Systems using Contrastive Adversarial Learning and Triplet Loss Fine-tuning

Though much work has been done in the domain of improving the adversarial\nrobustness of facial recognition systems, a surprisingly small percentage of it\nhas focused on self-supervised approaches. In this work, we present an approach\nthat combines Ad-versarial Pre-Training with Triplet Loss\nAdversarialFine-Tuning. We compare our methods with the pre-trained ResNet50\nmodel that forms the backbone of FaceNet, finetuned on our CelebA dataset.\nThrough comparing adversarial robustness achieved without adversarial training,\nwith triplet loss adversarial training, and our contrastive pre-training\ncombined with triplet loss adversarial fine-tuning, we find that our method\nachieves comparable results with far fewer epochs re-quired during fine-tuning.\nThis seems promising, increasing the training time for fine-tuning should yield\neven better results. In addition to this, a modified semi-supervised experiment\nwas conducted, which demonstrated the improvement of contrastive adversarial\ntraining with the introduction of small amounts of labels.\n