An investigation of privacy preservation in deep learning-based eye-tracking

Table 1 Validation loss (binary cross-entropy) scores for different inputs for the frame labeling network (instance, person)

	2 output^a	+2 grad^b	+5 grad^c	+2 grad + loss + label^d	+3 grad + loss + label^e
loss_Instance	0.7	0.61	0.61	0.57	0.57
loss_Person	0.7	0.61	0.62	0.59	0.58

For reference, the baseline (binary cross-entropy of random guesses for a balanced data) would be \(- \log(1/2) = 0.693\).
^aTakes the two last outputs (the target model output and the layer just before it) of the target model as the input
^bTakes the two last outputs and also the two gradients before the last gradient of the target model
^cTakes all the “+2 grad” and also the last gradient of three different sections of the target model (boundary, face, eyes)
^dTakes the two last layers outputs and also the two before the last gradient and label and loss of target model
^eTakes the two last layers outputs and also the three last gradients and label and loss of target model

ISSN: 1475-925X