Identity Independent Face Anti-spoofing Based on Random Scan Patterns.

Conventional face anti-spoofing paradigms tend to operate on plain facial profiles and learn either the natural face space alone (one-class training problem) or both the natural face space as well as the spoof sample space (2-class training problem). However, this rigidity with respect to spatially constrained measurements, makes the base feature or statistic vulnerable to noise related to pose and camera perspective/orientational and scale changes. Noting that the sharpness profile computed on a natural face is largely independent of the pose and perspective change, it is imperative that the measurements be extracted in an identity independent setting by ignoring the pose/perspective variation. To facilitate this, we have deployed a 2-dimensional random walk for capturing lower order pixel correlation statistics from natural faces, with virtually no perceptual interference. The proposed identity independent frame has surpassed the state of the art with reference to a 3D mask dataset (image oriented, isolated frame setting), with an EER of \(2.25\%\) without auto-population and an EER of \(0.45\%\) with auto-population.