Deep learning with spectrogram image of eye movement for biometrics

Biometric studies are being used worldwide for a large variety of purposes, here we used eye movements (EM) from observers of natural images. Since some EM are involuntary, these prevent spoofing attacks. While prior research requires feature extraction manually from EM data to identify a person, we use a deep convolutional architecture that processes it as an image. The eye movements were treated as a signal, then transformed as a spectrogram of frequencies, and its image is the input for a convolutional architecture. We investigated two types of signals: Cartesian coordinates, and gaze angle over time. The proposal consists of a convolutional network architecture applied to the DOVES dataset, where stimuli are natural images. We obtained the accuracy for the eye angle spectrogram, on DOVES, about 73%, and for the eye coordinates spectrogram, 65%. These results indicated that EM can be treated as spectrogram images for biometric identification.