Adversarial Attacks on Face Recognition Systems

Abstract

Face recognition has been widely used for identity verification both in supervised and unsupervised access control applications. The advancement in deep neural networks has opened up the possibility of scaling it to multiple applications. Despite the improvement in performance, deep network-based Face Recognition Systems (FRS) are not well prepared against adversarial attacks at the deployment level. The output performance of such FRS can be drastically impacted simply by changing the trained parameters, for instance, by changing the number of layers, subnetworks, loss and activation functions. This chapter will first demonstrate the impact on biometric performance using a publicly available face dataset. Further to this, this chapter will also present some strategies to defend against such attacks by incorporating defense mechanisms at the training level to mitigate the performance degradation. With the empirical evaluation of the deep FRS with and without a defense mechanism, we demonstrate the impact on biometric performance for the completeness of the chapter.