Eye Know You Too: A DenseNet Architecture for End-to-end Eye Movement Biometrics

TL;DR

The proposed DenseNet architecture for end-to-end EMB not only outperforms the previous state of the art, but is also the first to approach a level of authentication performance that would be acceptable for real-world use.

Abstract

Eye movement biometrics (EMB) is a relatively recent behavioral biometric modality that may have the potential to become the primary authentication method in virtual- and augmented-reality devices due to their emerging use of eye-tracking sensors to enable foveated rendering techniques. However, existing EMB models have yet to demonstrate levels of performance that would be acceptable for real-world use. Deep learning approaches to EMB have largely employed plain convolutional neural networks (CNNs), but there have been many milestone improvements to convolutional architectures over the years including residual networks (ResNets) and densely connected convolutional networks (DenseNets). The present study employs a DenseNet architecture for end-to-end EMB and compares the proposed model against the most relevant prior works. The proposed technique not only outperforms the previous state of the art, but is also the first to approach a level of authentication performance that would be acceptable for real-world use.

Figures (6)

• Figure 1: Overview of the process for embedding an eye-tracking signal using the proposed methodology. We primarily focus on the case where only the first 5-second window is embedded, but we explore aggregating embeddings across windows in § \ref{['sec:results-window']}.
• Figure 2: The proposed pre-activation densenet-based network architecture, including the optional classification layer. Each convolution layer has kernel size $k=3$, stride $s=1$, and dilation rate $d$ that varies by layer. Each convolution layer outputs 32 feature maps that are concatenated with the previous feature maps before being fed into the next convolution layer.
• Figure 3: (A) A visualization of the learning rate schedule used during training. (B) Progression of MAP@R (measured on TEX only) throughout training.
• Figure 4: Qualitative results for the primary evaluation setting: 5 s of R1 TEX. (A) Genuine and impostor similarity score distributions. (B) ROC curve for bootstrapped similarity score distributions (see § \ref{['sec:results-bootstrap']} for an explanation of how the bootstrapped distributions are made). The dashed black line shows where FRR and FAR are equal. The blue line is the mean ROC curve across 1000 bootstrapped distributions, and the shaded region represents $\pm$1 standard deviation around the mean.
• Figure 5: DensMAP narayan2020densmap visualizations of the embedding space for 10 subjects present across all rounds. All embeddings of valid ($\leq$50% NaNs) windows across all rounds R1--9 and both sessions are plotted together. A different mapping is fit for each plot. (A) Embeddings from only the TEX task. (B) Embeddings from all tasks (including BLG). We use umap-learn mcinnes2018umap-software parameters metric=cosine, n_neighbors=30, min_dist=0.1, and densmap=True.