Eye Know You: Metric Learning for End-to-end Biometric Authentication Using Eye Movements from a Longitudinal Dataset

TL;DR

The results reveal that reasonable authentication accuracy may be achieved even during both a low-cognitive-load task and at low sampling rates, and it is found that eye movements are quite resilient against template aging after as long as 3 years.

Abstract

The permanence of eye movements as a biometric modality remains largely unexplored in the literature. The present study addresses this limitation by evaluating a novel exponentially-dilated convolutional neural network for eye movement authentication using a recently proposed longitudinal dataset known as GazeBase. The network is trained using multi-similarity loss, which directly enables the enrollment and authentication of out-of-sample users. In addition, this study includes an exhaustive analysis of the effects of evaluating on various tasks and downsampling from 1000 Hz to several lower sampling rates. Our results reveal that reasonable authentication accuracy may be achieved even during both a low-cognitive-load task and at low sampling rates. Moreover, we find that eye movements are quite resilient against template aging after as long as 3 years.

Figures (5)

• Figure 1: Visualization of exponentially-dilated convolutions with kernel size 3, stride 1, and no padding. The convolutions in the $\ell$-th layer use a dilation of $d=2^{\ell-1}$. With this configuration, if the input has length $2^q$ and there are $q - 1$ layers, then the final layer has a receptive field of $2^q-1$ values from the input (shown in red and blue, with overlap in purple).
• Figure 2: Network architecture. Each convolution layer uses kernel size 3, stride 1, and no padding, and is followed by ReLU and batch normalization. The first fully-connected layer is followed by ReLU. The output of the final fully-connected layer acts as the embedding of the input. The numbers at the bottom reflect the shape of the data leaving each layer, ordered as minibatch size, channels, and time steps. The $d$ above each convolution block is the dilation used in that layer, and the $n$ above each fully-connected block is the number of output nodes in that layer. This network has a total of 475,264 learnable parameters.
• Figure 3: A representative comparison between empirical and resampled similarity distributions. EERs based on the resampled distributions tend to be slightly pessimistic relative to EERs based on the empirical distributions.
• Figure 4: Plots of the similarity distributions for genuine and impostor pairs. Each plot contains the similarities on the held-out test set computed separately for each of the 4 models trained with 4-fold cross-validation, using $n = 10$ for Equation \ref{['eq:sim']} and R1 for authentication. Since cosine similarity is bounded from -1 to 1, we scaled the similarities to lie between 0 and 1 before plotting. A bin width of 0.01 was used, and the area under each curve sums to 1.
• Figure 5: ROC curves to provide a qualitative assessment of model performance using $n = 10$ and R1 for authentication. The horizontal axis is log-scaled false acceptance rate (FAR). The vertical axis is false rejection rate (FRR). Each ROC curve represents the mean performance across 4 models trained with 4-fold cross-validation, and each shaded region indicates $\pm$1 SD about the mean. The point where the dashed line intersects each ROC curve indicates the EER for that curve.