Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Temporal Persistence and Intercorrelation of Embeddings Learned by an End-to-End Deep Learning Eye Movement-driven Biometrics Pipeline

Mehedi Hasan Raju, Lee Friedman, Dillon J Lohr, Oleg V Komogortsev

TL;DR

This paper investigates whether temporal persistence of embeddings in a DL-based eye-movement biometric pipeline is predictive of authentication performance. Using two public datasets (GB and GBVR) and four signal-quality manipulations (decimation, data-length, number of sequences, and Gaussian-noise-induced spatial-precision degradation), it analyzes how embedding reliability, intercorrelation, and $EER$ respond. Temporal persistence, measured by $KCC$, emerges as a strong predictor of biometric performance, with higher $KCC$ associated with lower $EER$ and embeddings typically weakly intercorrelated. The results support generalizing the temporal-persistence principle to end-to-end eye-movement biometrics and underscore the EKYT architecture as a robust framework, while noting limitations from small sample sizes and the need for replication.

Abstract

What qualities make a feature useful for biometric performance? In prior research, pre-dating the advent of deep learning (DL) approaches to biometric analysis, a strong relationship between temporal persistence, as indexed by the intraclass correlation coefficient (ICC), and biometric performance (Equal Error Rate, EER) was noted. More generally, the claim was made that good biometric performance resulted from a relatively large set of weakly intercorrelated features with high ICC. The present study aimed to determine whether the same relationships are found in a state-of-the-art DL-based eye movement biometric system (``Eye-Know-You-Too''), as applied to two publicly available eye movement datasets. To this end, we manipulate various aspects of eye-tracking signal quality, which produces variation in biometric performance, and relate that performance to the temporal persistence and intercorrelation of the resulting embeddings. Data quality indices were related to EER with either linear or logarithmic fits, and the resulting model R^2 was noted. As a general matter, we found that temporal persistence was an important predictor of DL-based biometric performance, and also that DL-learned embeddings were generally weakly intercorrelated.

Temporal Persistence and Intercorrelation of Embeddings Learned by an End-to-End Deep Learning Eye Movement-driven Biometrics Pipeline

TL;DR

This paper investigates whether temporal persistence of embeddings in a DL-based eye-movement biometric pipeline is predictive of authentication performance. Using two public datasets (GB and GBVR) and four signal-quality manipulations (decimation, data-length, number of sequences, and Gaussian-noise-induced spatial-precision degradation), it analyzes how embedding reliability, intercorrelation, and respond. Temporal persistence, measured by , emerges as a strong predictor of biometric performance, with higher associated with lower and embeddings typically weakly intercorrelated. The results support generalizing the temporal-persistence principle to end-to-end eye-movement biometrics and underscore the EKYT architecture as a robust framework, while noting limitations from small sample sizes and the need for replication.

Abstract

What qualities make a feature useful for biometric performance? In prior research, pre-dating the advent of deep learning (DL) approaches to biometric analysis, a strong relationship between temporal persistence, as indexed by the intraclass correlation coefficient (ICC), and biometric performance (Equal Error Rate, EER) was noted. More generally, the claim was made that good biometric performance resulted from a relatively large set of weakly intercorrelated features with high ICC. The present study aimed to determine whether the same relationships are found in a state-of-the-art DL-based eye movement biometric system (``Eye-Know-You-Too''), as applied to two publicly available eye movement datasets. To this end, we manipulate various aspects of eye-tracking signal quality, which produces variation in biometric performance, and relate that performance to the temporal persistence and intercorrelation of the resulting embeddings. Data quality indices were related to EER with either linear or logarithmic fits, and the resulting model R^2 was noted. As a general matter, we found that temporal persistence was an important predictor of DL-based biometric performance, and also that DL-learned embeddings were generally weakly intercorrelated.
Paper Structure (31 sections, 7 figures, 3 tables)

This paper contains 31 sections, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Prior Work
  3. Prior work on Temporal persistence/Biometric permanence
  4. Prior work on Eye Movement Biometrics
  5. Methodology
  6. Dataset
  7. Model Architecture and Data Handling
  8. Data Preprocessing
  9. Network Architecture
  10. Dataset Split & Training
  11. Embedding Generation
  12. Evaluation Metrics
  13. Hardware & Software
  14. Experiment Design
  15. RQ1: Decimation Do reductions in the sampling rate affect biometric performance, and are these changes related to the reliability of the learned embeddings?
  16. ...and 16 more sections

Figures (7)

  • Figure 1: Visual Representation of the Experimental Design. (A1) The interval between the red-dotted lines is defined as a sequence, containing 5000 samples for GB (for GBVR it is 1250 samples, not shown in the figure). (A2) Displays the first sequence from plot (A1). (B1) The signal from the plot (A1) has been downsampled to 25 Hz for demonstration. (B2) Shows the first sequence from plot (B1). (C1) Analyze only the first 10% of the signal, but place it in the center of the sequence with zero-padding on both sides. (C2) Presents the last sequence from the plot (C1) as an example. The right column provides a clearer visualization of the specific sequences from each row.
  • Figure 2: Relationship between KCC and EER with the decimated level (Hz) for two datasets: GB and GBVR. Subplots (A1) and (B1) show the logarithmic decrease in KCC for the GB and GBVR datasets, respectively, with lower sampling rates. Subplots (A2) and (B2) depict the logarithmic increase in EER for the GB and GBVR datasets, with lower sampling rates. Subplots (A3) and (B3) illustrate the strong negative logarithmic relationship between EER and KCC for the GB and GBVR datasets. $R^2$, and coefficient values are added to each plot's legend. The $R^2$ values across all plots indicate a strong fit, suggesting that a higher sampling rate improves biometric performance, with the GBVR dataset demonstrating a particularly robust fit.
  • Figure 3: Relationship between KCC and EER with the percentage level (%) for two datasets: GB and GBVR. Subplots (A1) and (B1) show the logarithmic decrease in KCC for the GB and GBVR datasets, respectively, with lower percentage levels. Subplots (A2) and (B2) depict the logarithmic increase in EER for the GB and GBVR datasets as the percentage levels decrease. The high $R^2$ values across these plots indicate a strong fit, suggesting that a higher percentage level improves biometric performance. Subplots (A3) and (B3) illustrate the strong negative logarithmic correlation between EER and KCC for the GB and GBVR datasets. $R^2$, and coefficient values are added to each plot's legend.
  • Figure 4: Relationship between the number of eye-tracking (ET) samples and the Equal Error Rate (EER). The plots illustrate the EER for two datasets, GB and GBVR, across varying sample sizes (50 to 5000 samples for GB and 50 to 1250 samples for GBVR). The results from RQ1 and RQ2 have been compared in each plot.
  • Figure 5: Relationship between KCC and EER with the number of sequences for two datasets: GB and GBVR. Subplots (A1) and (B1) show the logarithmic decrease in KCC GB and KCC GBVR, respectively, with reduced sequences. Subplots (A2) and (B2) depict the logarithmic increase in EER GB and EER GBVR with increasing sequences. Subplots (A3) and (B3) illustrate the strong negative logarithmic correlation between EER and KCC for GB and GBVR datasets, presenting higher temporal persistence associated with lower equal error rates. $R^2$, and coefficient values are added to each plot's legend.
  • ...and 2 more figures