Source-linked AI summary
Accuracy comparison across face recognition algorithms: Where are we on measuring race bias?
Jacqueline G. Cavazos, P. Jonathon Phillips, Carlos D. Castillo, Alice J. O'Toole
TL;DR
Race bias in face-recognition algorithms varies across data, algorithms, and application scenarios, making measurement a methodological challenge. The paper examines four algorithms across East Asian and Caucasian faces and finds that item difficulty, decision thresholds, and demographic composition affect observed bias and accuracy. It concludes that bias should be measured for individual applications, algorithms, races, and datasets.
Problem
Face-recognition algorithms have shown race-related accuracy differences, but assessing bias requires accounting for data-driven and scenario-based factors beyond training data alone.
Method
The paper reviews measurement factors and evaluates four algorithms, including one pre-DCNN system and three DCNNs, on East Asian and Caucasian faces.
Results
Accuracy improved across algorithm generations, but threshold-dependent differences remained, and race differences increased as item challenge level increased.
Takeaways & Limitations
Race bias must be measured for each particular scenario, algorithm, race, and dataset rather than assessed as a single general property.
Takeaways & Limitations
The study analyzed only two racial groups, and the dataset’s photometric and demographic control limited ecological validity.
Abstract
from arXiv · showhide
Previous generations of face recognition algorithms differ in accuracy for images of different races (race bias). Here, we present the possible underlying factors (data-driven and scenario modeling) and methodological considerations for assessing race bias in algorithms. We discuss data driven factors (e.g., image quality, image population statistics, and algorithm architecture), and scenario modeling factors that consider the role of the "user" of the algorithm (e.g., threshold decisions and demographic constraints). To illustrate how these issues apply, we present data from four face recognition algorithms (a previous-generation algorithm and three deep convolutional neural networks, DCNNs) for East Asian and Caucasian faces. First, dataset difficulty affected both overall recognition accuracy and race bias, such that race bias increased with item difficulty. Second, for all four algorithms, the degree of bias varied depending on the identification decision threshold. To achieve equal false accept rates (FARs), East Asian faces required higher identification thresholds than Caucasian faces, for all algorithms. Third, demographic constraints on the formulation of the distributions used in the test, impacted estimates of algorithm accuracy. We conclude that race bias needs to be measured for individual applications and we provide a checklist for measuring this bias in face recognition algorithms.
1. Introduction
The paper reviews race effects in human and computational face recognition and examines how measurement factors shape observed algorithmic race bias.
- The authors review prior human and algorithmic race effects before assessing race bias in newer face recognition systems.The study presents data from three DCNNs and one previous-generation algorithm across varying item difficulty.
- The analysis focuses on how item difficulty affects performance differences between East Asian and Caucasian faces.The authors argue that bias concerns become magnified as comparison items become more challenging.
- The paper combines methodological considerations with empirical measurements to evaluate race bias across algorithm generations.Its stated goal is to connect prior findings with present-day DCNN-based systems.
1.1. Other-race effect - Humans
Human face recognition shows an other-race effect: people recognize own-race faces relatively more accurately than other-race faces, with experience contributing to this pattern.
- The other-race effect is a statistical interaction showing relatively higher accuracy for own-race than other-race faces.This effect has been observed across multiple racial and ethnic groups and methodological paradigms.
- Face-recognition expertise develops partly through meaningful experience discriminating among faces across the lifespan.The cited work describes relevant experience beginning in infancy and continuing across age groups.
- Experience with homogeneous populations may produce specialized features that encode within-group facial differences effectively.These features are described as less suited to encoding the distinctive character of other-race faces.
- Human race bias reflects both an own-race accuracy advantage and the possibility that suitable features could make races equally recognizable.The latter conclusion is presented as an implication of replicated effects and experience-based findings.
1.2. Racial Bias - Algorithms
Across decades of research, face-recognition algorithms have shown race-related accuracy differences, with effects shaped by training experience, algorithm origin, demographics, and measurement conditions.
- 1.2.1. Pre-DCNN Algorithms.: Training-data composition influenced early algorithms, which recognized majority-race faces more accurately than minority-race faces.This pattern was attributed to greater model “experience” with the majority race.
- 1.2.1. Pre-DCNN Algorithms.: Algorithm origin was reported to mediate race bias, with a Western algorithm favoring Caucasian faces and an East Asian algorithm favoring East Asian faces.The compared systems were black-box algorithms, and training-set composition may have contributed to the pattern.
- 1.2.1. Pre-DCNN Algorithms.: Across commercial and in-house systems, young, Black, and female faces showed lower performance, while equitable training reduced but did not eliminate demographic biases.The reported findings involved race, gender, and age comparisons.
- 1.2.2. DCNN Algorithms.: DCNNs increased face-recognition accuracy and generalizability by using multilayer convolution and pooling operations to produce compressed face representations.These top-layer descriptors can be evaluated for race and other demographic effects.
- 1.2.2. DCNN Algorithms.: Race-bias findings for newer systems remain mixed across algorithms and scenarios, including differences in thresholds and skin-reflectance effects.Reported studies found algorithm-specific performance differences and lower similarity scores associated with darker skin reflectance.
- Nearly all studied face-recognition algorithms have shown performance differences as a function of face race.The review covers previous-generation systems and newer DCNNs, including evidence from NIST and other studies.
1.3. Measuring Face Identification Accuracy
Face-recognition accuracy can be summarized across all thresholds or at an application-specific threshold, and these measures may yield different conclusions about race bias.
- Signal Detection Theory frames verification as comparing similarity-score distributions for same-identity and different-identity image pairs.Higher similarity scores indicate greater likelihood that two images show the same person.
- ROC curves and AUC summarize overall discriminability, with AUC = 0.5 indicating chance and AUC = 1.0 indicating perfect accuracy.Lower AUC reflects greater distributional overlap and poorer discriminability.
- Identification requires a similarity threshold that determines false rejection rate and false accept rate.FRR concerns same-identity pairs judged different, whereas FAR concerns different-identity pairs judged the same.
- Applications often set thresholds to produce very small false-accept proportions, making verification rate at the user-set threshold practically important.Examples include 1/1,000 and 1/10,000 false-accept rates.
- Threshold-independent measures such as AUC and threshold-dependent measures such as VR @ FA=0.001 serve different purposes and may not converge.AUC can appear near ceiling while low-FAR verification rates still reveal meaningful race differences.
1.4. Factors underlying race bias
Race bias reflects both data-driven differences in face distributions or representations and scenario choices controlled by researchers or system users. Thresholds and demographic matching can change measured accuracy and bias across racial groups.
- Factors underlying race bias: Race-bias mechanisms fall into data-driven and scenario-modeling issues.Data-driven factors concern underlying distributions and representations; scenario-modeling factors concern thresholds and different-identity distributions.
- Data-driven factors: Representation quality may differ across demographic subgroups because of algorithm architecture or training data.Poor representations for particular subgroups can produce race bias across datasets.
- Data-driven factors: Test-image composition, image quality, subgroup structure, and item difficulty can alter estimated race bias.Image-quality differences explain some, but not all, variation; nested subgroups may amplify or hide effects.
- Scenario modeling: A uniform threshold may produce different FAR and VR estimates when subgroup distributions differ.Prior studies found that thresholds needed for consistent FAR or VR vary across racial groups.
- Scenario modeling: East Asian faces required higher thresholds than Caucasian faces to achieve equivalent FARs in the cited comparison.Threshold functions showed larger shifts for Asian faces, and this pattern was also observed across four algorithms for African American faces.
- Scenario modeling: Yoking controls demographic comparability in different-identity pairs, preventing demographic heterogeneity from artificially inflating verification performance.Accuracy decreased as demographic control increased in prior yoking comparisons.
- Implications: Race-bias assessments should evaluate threshold and yoking effects separately for each race of interest and each application.The paper treats these scenario choices as partially under researcher or system-user control.
2. Race Bias in Face Identification Algorithms
The experiment examines whether race bias varies with stimulus difficulty and how yoking changes estimated identification accuracy across older and newer algorithms.
- Experiment: Four face recognition algorithms were tested on Caucasian and East Asian face pairs across calibrated difficulty levels.The study also compares no yoking, race yoking, and race-and-gender yoking across algorithm generations.
2.1. Methods
The study uses selected indoor images from the Good, Bad, Ugly dataset and compares four algorithms on East Asian and Caucasian faces.
- Stimuli: The GBU dataset partitions image pairs into Good, Bad, and Ugly difficulty levels based on FRVT 2006 similarity scores.Each full partition contains 1,085 images of 437 identities.
- Stimuli: Indoor images were selected to reduce illumination variation, and analysis was limited to East Asian and Caucasian faces.These groups represented the majority of images in the selected dataset.
- Algorithms: The comparison included A2011, A2015, A2019, and A2017b, spanning a pre-DCNN algorithm and three DCNNs.A2011 was the oldest tested algorithm, while A2015 served as an established DCNN benchmark.
2.2. Results
Overall verification accuracy was summarized with ROC curves and AUC, while yoking conditions were used to examine how demographic matching changes performance estimates.
- Overall Accuracy: A2019 and A2017b showed the best overall accuracy, followed by A2011 and A2015.The ROC curves were plotted on a log scale around FAR = 0.0001 and 0.001.
- Yoking: Yoking comparisons used no demographic control, race control, and race-and-gender control for different-identity pairs.The corresponding figure distinguishes these conditions with dashed, solid, and dotted lines.
2.3. Race Bias
Race-bias estimates depend on both the evaluation metric and the operating threshold. Across four algorithms, low-FAR analyses revealed Caucasian advantages and equal FARs required higher thresholds for East Asian faces.
- ROC Curves: AUC was near ceiling and indicated little to no overall race bias, whereas low-FAR verification estimates revealed race bias for all algorithms.For A2019 and A2017b, bias appeared at FARs below 0.001 but not generally above that level.
- ROC Curves: At low FARs, all algorithms performed more accurately for Caucasian faces than for East Asian faces.The ROC curves show this comparison despite AUC detecting race bias for A2015 only.
- Thresholds: For all four algorithms, East Asian threshold functions shifted rightward relative to Caucasian functions.The shift means matching FARs cannot generally be achieved with one shared threshold.
- Thresholds: Threshold-shift magnitudes cannot be directly compared across algorithms because their similarity-score scales differ.The limitation concerns cross-algorithm comparisons of shift size, not the direction of the observed shifts.
- Thresholds: Equal FARs for East Asian and Caucasian faces required separate, higher thresholds for East Asian faces.A fixed threshold produces a smaller FAR for Caucasian faces than for East Asian faces.
2.4. Item Difficulty
Item difficulty affected both verification accuracy and race bias. Race disparities were generally more evident in harder image partitions than in nearly perfect, easier conditions.
- Item Difficulty: Accuracy for the Ugly partition was lowest across all three difficulty groups.A2019 and A2017b outperformed A2015 and A2011 across all partitions, while the lowest-performing algorithm varied by partition.
- Race Bias as a Function of Item Difficulty: In the Good partition, accuracy was nearly perfect for both East Asian and Caucasian faces.A2015 and, to a lesser extent, A2011 favored Caucasian faces at FAR = 0.0001.
- Race Bias as a Function of Item Difficulty: In the Bad partition, A2017b and A2019 showed nearly no race bias, while A2011 and A2015 favored Caucasian faces at FAR = 0.0001.The pattern varied across algorithms even at the same difficulty level.
- Race Bias as a Function of Item Difficulty: In the Ugly partition, every algorithm except A2011 was more accurate for Caucasian than East Asian faces.No algorithm achieved perfect performance for either race in this hardest partition.
3. Discussion
The paper argues that race bias in face recognition cannot be assessed with a single general measure because data-driven and scenario-modeling factors interact. Its empirical findings show that accuracy, bias, and conclusions about algorithm performance depend on item difficulty, algorithm generation, thresholds, and dataset design.
- Study scope: The study combines five lessons from prior human and algorithm research with new empirical data from three recent DCNN algorithms and one pre-DCNN algorithm.The empirical data examine race bias as a function of item-pair difficulty.
- Methodological considerations: Race bias must be measured for each scenario, algorithm, race, and dataset because interacting data-driven and scenario-modeling factors make general assessment unfeasible.Data-driven factors include population distributions, algorithm representations, representative images, and imaging conditions; scenario modeling includes threshold selection and demographic constraints.
- Item difficulty: As item challenge level increased, demographic differences were magnified, indicating that item difficulty affected estimates of race bias.The Ugly partition was the least accurate, and Figure 8 reports race bias for all algorithms except A2011 in that partition.
- Algorithm evolution: Accuracy improved for both race groups across algorithm generations, but threshold-dependent measures still showed problematic accuracy differences at specific operating points.The paper therefore treats overall accuracy gains and threshold-specific disparities as simultaneous findings.
- Limitations: The study was limited to East Asian and Caucasian groups, and the GBU dataset’s photometric and demographic controls may reduce ecological validity.The authors state that the methodological lessons apply across race groups despite the two-group analysis.
4. Conclusion
The paper concludes that assessing race and demographic bias requires a holistic approach covering both data-driven and scenario-modeling factors. Their interactions affect bias estimates, so simultaneous attention to all potential sources is needed for future studies.
- Conclusion: Both data-driven and scenario-modeling factors, including their interactions, can impact race bias in face recognition algorithms.The authors present this holistic assessment as a starting point for future studies.