Faking it, Making it: Fooling and Improving Brain-Based Authentication with Generative Adversarial Networks

In this paper, we empirically demonstrate the vulnerability of a passthought authentication system to fake signals generated by Generative Adversarial Networks (GANs), and use these same signals to make authenticators more robust. We first train a classifier that is able to authenticate a subject based on their EEG signals. The classifier performs a binary classification task: either the user is who they claim to be, or not. To test the robustness of the authenticator against attacks we train a GAN to generate signals that mimic the EEG signals of the “positive” subject. We find that a well-trained GAN is able to generate signals that the classifier consistently accepts. To alleviate this vulnerability, we re-train the classifier with this GAN-generated data. We find that the classifier re-trained against synthetic data is both more robust against this attack, and more accurate in accepting real data than the initial classifier. We conclude with recommendations for the design of passthought authentication systems.