BrainID: Using Twin Neural Networks for EEG-based Biometric Authentication

Abstract

Biometric authentication is any form of identity verification that uses a physical or physiological signal of a human, covering a wide variety of modalities ranging from iris patterns to DNA. Because most biometric factors are unique to the individual, difficult to spoof or steal, and do not require effortful retrieval, they are typically considered the most reliable form of identity verification. In recent years, deep learning and neural signal acquisition have seen tremendous advances, opening up the possibility for neural biometrics. We first address the challenge of neural biometric authentication from a variety of perspectives, including privacy, user-friendliness, resistance to attacks, universality, permanency, uniqueness, and collectability. Next, we summarize the state of the art in deep learning-based authentication systems, especially as they relate to EEG-specific applications. We then propose a twin neural network architecture to extract features which allow for differentiating EEG signals, even in the case of subjects which are not available during network training. We analyze the effectiveness of this technique as a neural biometric authentication method. We finish by assessing limitations of this analysis and future directions for research. The applications of this work include virtual and augmented reality device authentication, passive multi-device authentication, and providing context to related models used for neural signal processing.