A residual feature-based replay attack detection approach for brainprint biometric systems

Brainprint biometrics, as an emerging biometric technology, have recently gained increasing attention based on the assumption that each individual has unique memory and knowledge that are capable of providing distinctness from others. Like all other biometric methods, adversaries can also circumvent and compromise brainprint biometric systems, for example, by incorporating small-scale noises into the brainprint template to synthesize a faked input. To address this security vulnerability, we propose a novel replay detection approach by taking advantage of noise residual features to detect if the input is adversely modified and generated by adding noises onto a legitimate brainprint template. Specifically, the proposed approach consists of two separate stages: the identity recognition stage, which uses the convolutional neural network (CNN) to classify the input brainwaves and thus verify the identity of the user; and the replay detection stage, which uses the ensemble classifier to detect if the brainwave signals have been compromised and manipulated by using noise residual features. Experimental results show that the proposed approach can effectively detect the replay attacks to the brainprint biometric systems, while maintaining a rather high level of user identification accuracy.