Exploring the Risk Posed by Adversarial Attacks on Face Recognition with CNN

Face recognition models are gaining rising attention, especially CNN-based models. They are employed in multiple applications, such as security and surveillance. However, those models are vulnerable to adversarial attacks. This study investigates some adversarial attack strategies to evaluate the capability of the CNN-based face recognition models against adversarial attacks. In addition, some attack methods are employed in the same model to assess the abilities of those adversarial attack methods. Two CNN-based face recognition models are trained to test model accuracy and loss. Then, two adversarial attack methods are applied to attack the models. The first attack method is the Fast Gradient Sign Method (FGSM). As for the second attack method, it is the Random Gradient Method (RGM). This study demonstrates that ArcFace and CosFace models are vulnerable to Adversarial Attacks. Furthermore, the FGSM method has a better attack success rate in comparison with RGM to CNN-based face recognition models.