Abstract
Eye movements (EM) can be used to automatically differentiate between users (biometric). EM however are not only caused by user characteristics (endogenous) but also by stimulus properties (exogenous). Understanding the influences of exogenous factors allows for a better evaluation of EM-based identification systems (EMBIS) in order to improve their performance. We investigate how introduction of grouping in random Gabor field stimuli affects accuracy of EMBIS. Groups were introduced based on the principle of similarity by aligning the orientation (orientation-condition) or reducing the luminance (luminance-condition) of a random one-third of the Gabor elements. Three grouping levels were used: none, weak and strong. Luminance and orientation were matched such that group detection accuracy was 65% (weak) and 82% (strong). EM were recorded for eight participants, 120 trials each, during the initial 10s after stimulus onset. In order to ensure that the participants explored the stimulus, they were asked to make a key-press when a red marker appeared anywhere on the screen, randomly 10-13s after stimulus onset. A support vector machine (13 EM features) achieved an overall user identification accuracy of 78.23%. No grouping resulted in lowest accuracy (70.50%), weak grouping in highest (78.30%) and strong grouping achieved 72.88%. Notably, grouping by luminance lead to lower accuracy (74.63%) than by orientation (77.93%). In conclusion, some exogenous properties must be present to invoke user characteristic EM (reflected in higher accuracy for weak, but lower accuracy for no grouping). Although, above a threshold, exogenous properties become a detrimental factor on EM based user identification (lower accuracy for strong than for weak grouping). The differences in accuracy for different cues, e.g., luminance and orientation, can be used to modulate activity of automated EMBIS. We therefore recommend further systematic investigation of exogenous factors with respect to their influence on accuracy of EMBIS.
Meeting abstract presented at VSS 2015