Research pick: Eye, eye - "Deep learning with spectrogram image of eye movement for biometrics"

Research in the International Journal of Biometrics has looked at how deep learning can be used to analyse spectrogram images of the human eye and its movements as a biometric tool.

Antonio Ricardo Alexandre Brasil and Patrick Marques Ciarelli of the Federal University of Espírito Santo in Vitória and Izabella Martins da Costa Rodrigues, Jefferson Oliveira Andrade, and Karin Satie Komati of the Federal Institute of Espírito Santo in Serra, Brazil, have developed a novel approach to personal identification based on eye movements for recognition and security applications. Their biometric technique has proven resilient to fraudulent attempts because it focuses on the involuntary nature of certain eye movements.

Conventionally, identifying individuals through eye movements required manual feature extraction from the data. Brasil and colleagues have circumvented the inherent problems with that approach by converting recorded eye movements into a data signal that can be processed by an algorithm trained on known data. Specifically, they use deep convolutional architecture to process the Cartesian coordinates, the points where the eyes are looking over time and the gaze angles of volunteers.

The team were able to achieve an accuracy of around 73% for eye angle spectrogram identification and 65% for eye coordinate spectrogram identification testing against the DOVES dataset. Taken together, this would be effective at identifying an individual from their unique pattern of eye movements and is, the team says, the first time spectrograms have been used in this way.

The spectrograms generated from gaze angles outperformed those based on only Cartesian coordinates. Future research using larger and more diverse datasets has the potential to improve analysis and accuracy still further, the team suggests. The researchers also plan to investigate alternative methodologies, such as using long short-term memory (LSTM) layers and fixation density map (FDM) to boost accuracy.

Brasil, A.R.A., Ciarelli, P.M., Rodrigues, I.M.d.C., Andrade, J.O. and Komati, K.S. (2023) ‘Deep learning with spectrogram image of eye movement for biometrics’, Int. J. Biometrics, Vol. 15, No. 6, pp.726–744.