Source-linked AI summary
Facial Age Estimation for Age Fraud Detection in National ID Systems
Sharib Athar, Arka Koner, Chetan Naik, Barada P. Sabut, Tanusree Deb Barma, Anoop M. Namboodiri, Anil K. Jain
TL;DR
Age misrepresentation can affect access to age-dependent services, while facial age estimation is complicated by aging variation, label noise, and capture conditions. SwinAge applies facial age estimation and landmark-based alignment to Aadhaar enrollment data, using threshold-based triage for review. It reports a 2.94-year MAE on its held-out test set and competitive performance on public benchmarks after fine-tuning.
Problem
Age-related misrepresentation can affect eligibility for age-dependent services, and facial age estimation faces variation in aging, label noise, and image capture conditions.
Source: §1 · pp. 1–2Method
SwinAge applies a SwinFace-based age estimator with landmark alignment to Aadhaar data and compares estimated and declared ages at operational thresholds.
Source: §1 · p. 2, §3 · p. 3, §5.2 · p. 5Results
2.94 years MAE on a 283K held-out test set, with competitive performance on public benchmarks after fine-tuning.
Source: §8 · p. 10Takeaways & Limitations
Threshold-based evaluation matters because aggregate MAE does not predict behavior at an eligibility threshold.
Source: §7.1 · p. 9SwinAge does not detect photo substitution, and robustness to appearance manipulation and degraded capture is not designed in.
Source: §7 · p. 9
Abstract
from arXiv · showhide
Identity fraud during biometric enrollment and updates remains a major challenge for large-scale national identity systems. A common fraud vector is misrepresenting one's age to access age-restricted services or welfare schemes. In this work, we present SwinAge, a facial age estimation system designed for use within the Aadhaar biometric enrollment pipeline, to assist quality-check (QC) operators to flag potential age-related fraud. This is critical for a system like Aadhaar (the world's largest national identity programme), that holds about 1.5 billion unique identities, with 22.4 million new enrollments and 283 million updates in the last year. Building upon the SwinFace architecture with landmark-based similarity (warp affine) alignment, we train on a large in-house dataset of 1.45 million face images and evaluate on an independent, age-stratified test set of 283K images, both drawn from an ethnically diverse population of 716K unique subjects. We investigate three Aadhaar-specific operational thresholds (5, 18, and 60 years) and propose a deployment triage framework that flags suspected cases for manual review. Following NIST FATE, we report false acceptance/rejection rates (FAR/FRR) at each threshold rather than aggregate accuracy: at 1% FAR the model achieves an FRR of 3% (<5yrs), 0.4% (>18yrs) and 11.0% (>60yrs). SwinAge achieves a mean absolute error (MAE) of 2.94 years on the same test set, outperforming three zero-shot vision language models on all benchmarks, and improving the state-of-the-art on 5 out of 7 public benchmark datasets. We further report per-gender errors and distill lessons for national identity programs.
1. Introduction
Age misrepresentation can affect eligibility for services and benefits in Aadhaar, while estimating age from faces is complicated by varied aging, data imbalance, label noise, and capture conditions. SwinAge is introduced to help flag substantial declared-versus-estimated age discrepancies for operator review.
1. Introduction
- More than Rs. 610 million was improperly disbursed to over 95,000 ineligible beneficiaries in a national audit, including people below age thresholds for pensions and disability benefits.The reported cases included 57,394 people under 60 receiving old-age pensions and 5,380 people under 18 receiving disability benefits.Source: §1 · p. 1
- SwinAge combines facial age estimation with declared-age comparison to flag cases for operator review, rather than replace document-based verification.The paper motivates this triage because manual verification is subjective and exhaustive checks are infeasible at national scale.Source: §1 · p. 2
- Age-related fraud includes claims of being under 5 to avoid biometric capture, being at least 18 for adult eligibility, or at least 60 for senior benefits.Children under 5 are enrolled with a photograph only; fingerprints and iris are captured at later mandatory updates.Source: §1 · p. 2
- Facial age estimation is challenged by non-linear aging, differences between people of the same age, sparse age extremes, noisy labels, and capture variability.NIST reports that accuracy also depends strongly on sex, image quality, region of birth, and age.Source: §1 · p. 2
- SwinAge is built on SwinFace and evaluated on Aadhaar enrollment data, public benchmarks, and three age-threshold fraud scenarios.The study reports analysis by age group and gender and examines alignment and backbone fine-tuning.Source: §1 · p. 2
2. Related Work
Prior facial age estimation work spans regression, classification, label-distribution learning, ordinal regression, and transformer-based approaches. This section also situates SwinAge among evaluation and demographic-bias studies, and highlights its national-ID data setting.
2. Related Work
- Age-estimation methods include regression, classification, label-distribution learning, ordinal regression, and hybrids such as mean–variance loss.DEX treats apparent-age estimation as classification followed by softmax expectation using VGG-16 ensembles.Source: §2 · p. 2
- Transformer approaches include SwinFace, MiVOLO, and FPAge, which use shared multi-task features, facial and full-body inputs, and face parsing, respectively.FPAge uses semantic face parsing with attention to focus on informative facial components under pose, expression, and occlusion variation.Source: §2 · pp. 2–3
- Prior work studies preprocessing and loss design, while reported gender and race bias motivates SwinAge’s per-gender and per-age-group error reporting.The cited work includes alignment, resolution, coverage, and pretraining, gender features with ordinal loss, and adaptive label-distribution learning [24].Source: §2 · p. 3
- The paper evaluates operational threshold error rates and notes that, to its knowledge, prior work had not tested age estimation on national-ID enrollment data where the reference label is the record under audit.The evaluation is framed against NIST FATE and ISO/IEC 27566-1.Source: §2 · p. 3
3. The SwinAge Model
SwinAge uses a SwinFace-based age-estimation head and landmark alignment, with a training loss that shifts toward robustness to noisy-label outliers. Its configuration keeps the pretrained backbone frozen except in an ablation.
3. The SwinAge Model
- SwinAge uses SwinFace multi-scale features with a Feature Attention Module, task-specific subnet, and age-regression output; the backbone is normally frozen.The model has approximately 36.0M parameters, of which 7.1M are trainable.Source: §3 · p. 3
3.1. Face Detection and Alignment
- Five-point landmark similarity alignment maps faces to a canonical 112 × 112 template, correcting in-plane rotation, translation, and scale before the backbone.The alignment is intended to match the backbone’s pretraining alignment; side-facing poses can be more distorted because the transform assumes a near-frontal face.Source: §3.1 · p. 3, Figure 3 · p. 3, Figure 4 · p. 4
3.2. Loss Function
- The hybrid age-prediction loss transitions from squared error toward a robust Gaussian loss to reduce the influence of large noisy-label outliers.Its progress ratio is λ = t/T, and the robust-regime scale σ is fixed at 3.0 based on validation.Source: §3.2 · pp. 3–4
3.3. Training Configuration
- Training starts from pretrained SwinFace weights and updates only the task-specific head, using AdamW, augmentation, and 80,000 training steps.Training takes approximately 60 hours on two Tesla V100-32GB GPUs using FP16 mixed precision.Source: §3.3 · p. 4
4. Datasets
The in-house dataset consists of anonymized Aadhaar enrollment images from a predominantly South Asian population, divided into subject-disjoint training, validation, and age-stratified test sets. Public benchmarks add evaluation across datasets with varied sizes, age ranges, image quality, and capture conditions.
4.1. In-House Dataset
- The in-house corpus contains 1.75 million images from 716,846 subjects and is described as diverse in skin tone, facial features, hairstyles, and aging patterns.Personally identifiable information was removed, and raw database images are not published or shared.Source: §4.1 · p. 4
- Subject-level splitting yields 1,451,294 training images, 16,941 validation images, and 283,515 test images, with no identity leakage between splits.The age-stratified test set covers 192,096 subjects; its overall MAE reflects a balanced rather than real-world age distribution, and labels are declared dates of birth rather than verified chronological ages.Source: §4.1 · p. 4
4.2. Public Datasets
- The evaluation includes six widely used public age benchmarks plus AFAD for within-dataset comparison, with combined-data experiments also described.Benchmark datasets differ in size, age range, image quality, and capture conditions; for combined training, datasets over 20,000 training samples are subsampled to limit dominance.Source: §4.2 · p. 4, Table 1 · p. 4
5. Experiments and Results
The experiments assess threshold-based fraud triage, architecture and alignment choices, demographic error patterns, and performance across public datasets. SwinAge achieves a 2.94-year in-house MAE, outperforms zero-shot vision-language models on every dataset, and matches or improves on five of seven public datasets.
5.1. Experimental Setup
- The evaluation compares seven configurations, including pretrained SwinFace, FP-Age variants, and SwinFace variants with different alignment methods.Models 2–7 use the same 1.45M-image training set and identical training configurations, enabling controlled comparison of architecture and preprocessing.Source: §5.1 · pp. 4–5
5.2. Use-Case-Driven Evaluation
- The threshold evaluation uses continuous predicted and true ages and reports FAR and FRR for confident errors under decision bands.The study evaluates thresholds 5, 18, and 60; the bands defer near-threshold cases to document verification, and the 1% FAR setting is read off the test set rather than independently validated.Source: §5.2 · p. 5
- At operational thresholds, the ≥60 FRR remains the binding deployment constraint across band settings, while FAR stays low at all thresholds.The exact rule rejects 21.4% of genuine seniors; reported band-setting rates count confident errors only.Source: §5.2 · pp. 5–6, Table 2 · p. 6
5.3.1. Ablation Study: Architecture and Alignment
- 2.94 years, SwinAge achieves the best overall MAE among the tested architecture and alignment configurations.Models 2–7 share training settings and vary only in architecture and alignment; the pretrained-weights baseline is not retrained.Source: §5.1 · p. 5, Table 3 · p. 6, §5.3.1 · p. 6
5.3.2. Per-Age-Group and Gender Analysis
- Male subjects have lower MAE than female subjects (2.79 versus 3.10), with the largest differences in ages 15–20; 42 transgender images from 26 subjects are too few for separate reporting.The authors recommend auditing demographic parity at the decision level, with ≥60 FRR under the most scrutiny.Source: §5.3.2 · p. 6, Figure 6 · p. 6
5.4. Public Dataset Evaluation
- Joint finetuning remains competitive on every dataset but performs worse than dataset-specific finetuning on each, including the in-house set.AFAD is excluded from combined “All” training; the authors attribute the pattern likely to domain conflict between public datasets and controlled in-house capture.Source: §5.4 · pp. 6–7, Table 4 · p. 7
- SwinAge outperforms all three zero-shot vision-language models on every dataset and matches or improves on five of seven public datasets.On in-house data, the comparison is 2.94 versus 4.16–5.14 MAE; protocol differences can shift specialized-estimator MAE by more than the gaps being compared.Source: §5.4 · p. 6, Table 4 · p. 7
6. Discussion
The discussion finds that alignment choice materially affects age-estimation error, while performance varies by age and dataset. It also documents regression-to-the-mean bias and the influence of outliers, and describes a qualitative review of difficult predictions.
6.1. Effect of Face Alignment
- Warp affine improves SwinAge MAE to 2.94 from 3.77 without alignment; for FPAge, tanh-polar slightly outperforms warp affine, with MAEs of 3.06 and 3.16.For controlled, near-frontal captures, SwinAge tanh-polar yields 3.88 MAE, worse than the 3.77 without alignment.Source: §6.1 · p. 7
- Unfreezing the backbone improves test MAE by only 0.03 years (2.91 vs. 2.94), with validation performance plateauing at step 10,000.The plateau persisted through the remaining 70,000 training steps.Source: §6.1 · p. 7
6.2. Age-Group-Specific Analysis
- Children aged 0–5 are easiest to estimate (MAE 0.75), while ages 45–70+ are hardest (MAE ∼3.75–4.44).The discussion attributes the hardest range to subtle, gradual aging changes and increasing differences among people of the same age; the corpus is capped at 75.Source: §6.2 · p. 7
6.2.1. Regression-to-Mean: A Per-Dataset Bias Analysis
- UTKFace and AgeDB show opposing age-bin biases despite near-perfect child estimates on UTKFace and near-zero overall bias on AgeDB.UTKFace has MAE 0.79 for ages 0–5 and −5.0 years of bias for ages 70+; AgeDB shifts from +3.3 to −4.8 years despite overall bias of −0.08.Source: §6.2.1 · p. 7
- Across datasets, predictions overestimate younger subjects and underestimate older ones, with a crossover near ages 30–40.Figure 7 plots mean prediction error by true age in 5-year bins; its broader <20 and ≥60 bins differ from the narrower bins discussed elsewhere.Source: §6.2.1 · p. 7, Figure 7 · p. 8
- Predicted-age variance compression is reported as 90–96% of true-age variance, with National ID at 97.3%, MORPH2 at 97.6%, and AgeDB at 82.4%.The discussion links this compression to regression-to-the-mean bias and notes that AgeDB’s wider age span amplifies edge effects.Source: §6.2.1 · p. 8
- FGNet’s 60–65 bin contains one heavily mislabeled image with MAE 14.1; its 82-subject sample makes results sensitive to individual outliers.Excluding that bin lowers variance compression to 95.2%.Source: §6.2.1 · p. 8
6.3. Qualitative Failure Analysis
- Figure 8 presents the two worst absolute-error predictions per public dataset; National ID images are omitted for privacy reasons.The examples are from the Intra setting.Source: §6.3 · p. 8
7. Lessons for National Identity Programmes
SwinAge’s operational trade-offs vary by age threshold and review volume. The lessons emphasize explicit decision rules, label provenance, subgroup sampling, and audit trails.
7. Lessons for National Identity Programmes
- Logging each discrepancy creates an audit trail, and unusually high operator- or centre-level discrepancy rates can trigger targeted audits.Source: §7 · p. 8
- Worst public-benchmark predictions involved extreme age with atypical appearance, celebrity apparent-age bias, mislabelling or detection errors, and degraded image quality or pose.Source: Figure 8 · p. 9
- MAE was 0.75 for ages 0–5, 2.07 for ages 15–20, and ∼3.76 for ages 55–65; senior-age error was largest and drove elevated ≥60 FRR.Source: §7 · p. 9
- At ±10-year discrepancy, 1.5% (4,126 of 283,515) were flagged; ±15 years flagged 0.2% (629), versus 19.5% (55,214) at ±5 years.Source: §7 · p. 9
7.1. Interpreting the Results
- MAE 2.94 coexisted with FRR 21.4% at τ=60 under the exact rule and 11.0% with a band; report performance at a stated operating point.Source: §7.1 · p. 9
- Margins read off test data are non-comparable and upper-bound operational performance; δ=0.1τ is a fallback, not a guarantee of target FAR.Source: §7.1 · p. 9
- A model trained on declared birth dates measures agreement with those records, not chronological age; provenance stratification and longitudinal self-consistency can help assess label quality.Source: §7.1 · p. 9
- A 0.3-year gender MAE gap does not characterize decision-level harm; subgroup audits require adequate samples, and 42 transgender images (26 subjects) were too few to report.Source: §7.1 · p. 10
8. Conclusion
The conclusion presents SwinAge as an age-fraud triage aid for Aadhaar, with held-out performance and decision-support safeguards. It also identifies senior-age estimation and label-noise mitigation as future-work priorities.
8. Conclusion
- SwinAge achieved MAE 2.94 on 283K held-out images; with the calibrated band, FAR was ≤1% at all thresholds and FRRs were 3.0%, 0.4%, and 11.0%.Source: §8 · p. 10
- The 60+ threshold remains most challenging, attributed to elevated MAE in that age range and the training corpus’s 75-year ceiling.Source: §8 · p. 10
- A flagged resident is told manual verification is required, may contest it before a quality check supervisor, and enrollment proceeds pending review; no benefit is withheld on a flag alone.Source: §8 · p. 10
- Facial images were accessed only for training and testing under controlled, on-premises conditions; the paper calls for regular audits of decision bias across age and skin tone.Source: §8 · p. 10
- Future work includes age progression as an auxiliary representation-learning task using 2.17 million paired same-subject images, and label-noise mitigation.Source: §8 · p. 10