How Many Persistent Features Does a Biometric Need? A Law Linking Equal Error Rate to Feature Count and Persistence
Organizations: Texas State University, San Marcos, Texas, USA
Abstract
A biometric designer controls the number of features and their temporal persistence. Earlier empirical work linked the equal error rate (EER) to persistence and to the required number of features by straight lines. We derive the law behind these regularities. For independent standardized Gaussian features with test-retest correlation , scored by cosine similarity, Fisher's -transform turns the genuine and impostor scores into two normals, giving and, inversely, with ; intercorrelated features enter through an effective count . The central extension lets persistence vary across comparisons with spread on Fisher's scale: with , which for a normal spread has a floor that no number of features can lower. With no fitted parameter the simple law reproduces published simulations and the EER-ICC line, and predicts the KCC-EER relation of eye-movement embeddings on GazeBase (), though less well on GazeBaseVR (). Read from the genuine and impostor scores, the general form matches the EER of eye movements, keystroke dynamics, the ECG and face embeddings over 101 conditions with and a median error of points, against and points for the simple law. The 61 eye-movement, keystroke and ECG conditions check consistency on the same scores; the 40 face conditions use held-out subjects, with the same accuracy. The product gives a design rule: adding or decorrelating features pays fully while it is well below one and little once it exceeds about three, where only higher or more uniform persistence helps.
Figures & tables
| Index | Relation to | Invariant to |
|---|---|---|
| Pearson | linear change | |
| if no session effect | nothing | |
| Spearman | monotone change | |
| Kendall | monotone change |
| Features | Probe | EER | law | two-normal | mixture |
|---|---|---|---|---|---|
| raw | session 2 | 5.88 | 5.17 | 5.87 | 5.76 |
| raw | session 8 | 17.29 | 17.99 | 18.33 | 19.44 |
| rank-normalized | session 2 | 7.96 | 8.44 | 8.27 | 5.98 |
| rank-normalized | session 8 | 18.71 | 23.42 | 21.16 | 22.32 |
| all pairs | held-out halves, 512 dimensions | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Gap (years) | mean gap | pairs | KCC | ICC | a-priori law | EER | EER | normal spread | measured spread | ( 18 ) | ||
| 3–5 | 4.0 | 566 | 0.756 | 0.528 | 0.002 | 3.18 | 0.19 | 3.03 | 0.00 | 1.04 | 2.69 | 2.90 |
| 6–10 | 8.0 | 565 | 0.743 | 0.501 | 0.006 | 3.89 | 0.20 | 3.68 | 0.00 | 1.65 | 3.32 | 3.73 |
| 11–20 | 15.3 | 563 | 0.713 | 0.438 | 0.05 | 6.04 | 0.19 | 5.66 | 0.01 | 3.10 | 5.04 | 5.57 |
| 21–30 | 25.4 | 553 | 0.692 | 0.395 | 0.15 | 5.97 | 0.17 | 6.05 | 0.02 | 3.51 | 4.93 | 5.64 |
| 31+ | 37.6 | 517 | 0.673 | 0.356 | 0.40 | 6.19 | 0.15 | 6.34 | 0.07 | 3.96 | 5.35 | 6.15 |
| median error (points) | ||||||||||
| variation | two- | normal | measured | |||||||
| Modality | points | of | limited by | normal | spread | spread | ||||
| Keystroke | 4 | little | 0.08 | 0.45 | 0.18 | features | 1.31 | 0.67 | 1.79 | 1.79 |
| Eye movements | 56 | modest | 0.21 | 0.26 | 0.83 | both | 1.12 | 0.63 | 0.63 | 0.55 |
| ECG | 1 | between people | 0.71 | 0.38 | 1.88 | mostly spread | 10.71 | 1.58 | 1.58 | 0.88 |
| Faces (held-out) | 40 | between photos | 0.19 | 0.06 | 2.93 | near spread limit | 3.69 | — | 1.62 | 0.59 |
| KCC ( ) | independent | |||
|---|---|---|---|---|
| 0.7 (0.42) | 50 | 6.1 | 10.3 ( ) | 21.8 ( ) |
| 100 | 1.4 | 6.0 ( ) | 20.1 ( ) | |
| 200 | 0.09 | 3.6 ( ) | 19.2 ( ) | |
| 0.8 (0.62) | 50 | 0.58 | 2.0 | 10.2 |
| 100 | 0.016 | 0.57 | 8.6 | |
| 200 | 0.17 | 7.8 |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| ICC | simulated | exact | law | law, | |
|---|---|---|---|---|---|
| 2 | 0.75 | 28.5 | 28.3 | 24.6 | 31.3 |
| 2 | 0.95 | 16.5 | 16.3 | 9.8 | 18.0 |
| 4 | 0.75 | 18.9 | 18.8 | 16.5 | 20.0 |
| 4 | 0.95 | 5.8 | 5.8 | 3.3 | 5.6 |
| 8 | 0.55 | 20.0 | 20.1 | 19.1 | 20.7 |
| 8 | 0.85 | 5.0 | 4.9 | 3.8 | 4.8 |
| ICC | 5% | 2% | 1% | 0.5% | 0.1% |
|---|---|---|---|---|---|
| 0.35 | 82 (82) [4.89] | 127 (127) [2.00] | 162 (163) [1.07] | 198 (200) [0.57] | 281 (287) [0.10] |
| 0.45 | 48 (47) [4.83] | 74 (73) [1.95] | 94 (93) [0.97] | 115 (114) [0.49] | 166 (164) [0.13] |
| 0.55 | 30 (29) [4.79] | 46 (45) [1.98] | 59 (58) [0.96] | 72 (70) [0.49] | 102 (101) [0.10] |
| 0.65 | 19 (19) [5.00] | 30 (29) [1.94] | 38 (37) [1.06] | 46 (45) [0.44] | 66 (65) [0.09] |
| 0.75 | 13 (12) [4.63] | 20 (19) [1.85] | 25 (24) [0.94] | 30 (29) [0.45] | 43 (41) [0.10] |
| 0.85 | 8 (8) [4.83] | 12 (12) [2.14] | 16 (15) [0.77] | 19 (18) [0.44] | 27 (25) [0.09] |
| ICC distribution | mean | med. | sim. | law(mean) | law(med.) |
|---|---|---|---|---|---|
| uniform 0.3–0.9 | 0.60 | 0.60 | 3.08 | 2.68 | 2.68 |
| uniform 0.1–0.5 | 0.30 | 0.30 | 19.4 | 19.4 | 19.4 |
| 75% at 0.3, 25% at 0.9 | 0.45 | 0.30 | 9.12 | 8.86 | 19.4 |
| 25% at 0.3, 75% at 0.9 | 0.75 | 0.90 | 0.70 | 0.34 | 0.002 |
| half 0.05, half 0.8 | 0.425 | 0.425 | 11.4 | 10.3 | 10.3 |
| same, | 0.425 | 0.425 | 4.58 | 3.59 | 3.59 |
| mean | SD | sim. | law | 2-norm. | mixt. | ||
|---|---|---|---|---|---|---|---|
| normal | 0.60 | 0.00 | 2.59 | 2.68 | 2.60 | 2.68 | 32 |
| normal | 0.60 | 0.05 | 3.48 | 2.68 | 3.30 | 3.17 | 28 |
| normal | 0.60 | 0.10 | 4.70 | 2.75 | 4.92 | 4.53 | 25 |
| normal | 0.60 | 0.15 | 6.08 | 2.77 | 8.13 | 6.21 | 21 |
| normal | 0.60 | 0.19 | 8.43 | 2.79 | 11.66 | 8.00 | 17 |
| normal | 0.60 | 0.23 | 10.00 | 2.78 | 13.41 | 9.84 | 15 |
| Pearson | law | two- | ||||||
|---|---|---|---|---|---|---|---|---|
| Change applied to session 2 | EER | KCC | law( ) | law(ICC) | ( ) | normal | ||
| none | 0.80 | 0.70 | 0.70 | 0.84 | 0.80 | 0.80 | 0.74 | 0.78 |
| shift SD, every dimension | 6.40 | 0.70 | 0.47 | 0.84 | 0.79 | 7.93 | 6.55 | 6.47 |
| shift random per dimension (SD 1) | 8.10 | 0.70 | 0.48 | 0.84 | 0.80 | 7.11 | 8.05 | 7.74 |
| uniform rescale | 0.84 | 0.70 | 0.42 | 0.84 | 0.78 | 10.56 | 0.73 | 0.76 |
| random rescale per dimension | 2.73 | 0.71 | 0.60 | 0.84 | 0.72 | 2.67 | 2.60 | 2.34 |
| Index over the 128 dimensions | corr. | mean off. | max off. |
|---|---|---|---|
| mean Pearson | 0.992 | 0.093 | |
| mean | 0.993 | 0.094 | |
| KCC, mapped by (12) | 0.993 | 0.107 | |
| variance-weighted Pearson | 0.994 | 0.093 |
| Condition | measured | two-norm. | general | ||
|---|---|---|---|---|---|
| GazeBase, unmanip. | 1.10 | 0.44 | 0.21 | 1.07 | 1.03 |
| GazeBaseVR, unmanip. | 0.68 | 3.33 | 2.29 | 3.52 | 3.47 |
| Keystroke raw, S2 | 0.00 | 5.88 | 7.16 | 5.87 | 7.16 |
| Keystroke raw, S8 | 0.00 | 17.29 | 19.59 | 18.33 | 19.59 |
| Keystroke rank, S2 | 0.36 | 7.96 | 7.62 | 8.27 | 8.24 |
| Keystroke rank, S8 | 0.45 | 18.71 | 20.05 | 21.16 | 21.16 |
| Diagnostic | Form | Example |
|---|---|---|
| (heavy-tailed features) | rank-normalize first (Section V-G); otherwise the two-normal form | raw keystroke latencies ( – ) |
| ( ) | the law (8) at and | rank-normalized keystroke ( – ) |
| ; normal and measured spread agree | closed form (15), equivalently the two-normal form | eye-movement embeddings ( ) |
| ; they disagree (heavy lower tail) | mixture (14): from three or more sessions if the spread is between people, deconvolved pilot genuine scores if between recordings | ECG ( ), faces ( – ) |
Biometric Identification
Biometric identification matches or verifies people and animals using distinctive physical, physiological, or behavioral signals.
See all 136 papers