Persona Selfie Age Estimation

---

Introduction

Understanding a user’s age is a critical requirement for many digital services, particularly those involving age-restricted goods, youth-focused platforms, or sensitive online communities. While age assessment has long existed in physical environments, the growth of fully online services introduces new challenges: service providers must determine age remotely, without face-to-face interaction or guaranteed access to government-issued identification.

Age estimation has emerged as an important component of modern digital identity systems. When combined with liveness detection, it enables automated, scalable age assurance that mirrors real-world practices such as conditional ID checks. This article describes Persona’s proprietary Selfie Age Estimation technology, including its ethical development, technical operation, and performance characteristics across multiple independent evaluations.

---

Terminology

Term	Description
Age Estimation Technology	Software which assesses images and determines the estimated age of faces in the field of view. Age estimates are characterized in terms of accuracy by Mean Absolute Error (MAE) and are typically reported as a certain number of years old.
Image Quality Metrics	Software which assesses images and determines if they are of sufficiently high quality to perform a given analysis (in this case, age estimation). If an image is of low quality (for example, low resolution, poor lighting, or blur), it may not be possible to accurately estimate the age of faces in the field of view.
Mean Absolute Error (MAE)	The primary metric for reporting the accuracy of age estimation software. MAE is calculated by determining the absolute error for each sample—how many years older or younger the estimated age is compared to the actual age—and then computing the mean across a statistically significant sample set. MAE is reported in years and represents the average expected error. Because absolute error is used, positive and negative errors do not cancel each other out. MAE represents an average, not a maximum; individual samples may have errors above or below the reported value.
False Positive Rate (FPR)	A measure of how frequently an age estimation system reports that a person who is actually of age X or lower is of age T or higher. For example, FPR may describe the likelihood that a person aged 15 or younger is estimated as being 18 or older. In this context, a false positive refers to an estimated age that exceeds the enforced threshold. In a “Challenge 30” model, a false positive would be incorrectly judging that a person under 21 appears to be over 30.
False Negative Rate (FNR)	A measure of how frequently an age estimation system reports that a person who is actually of age T or higher is of age X or lower. For example, FNR may describe the likelihood that a person aged 18 or older is estimated as being 15 or younger. In this context, a false negative refers to an estimated age that falls below the enforced threshold. In a “Challenge 30” model, a false negative would be incorrectly judging that a person over 30 appears to be under 21.

---

Ethical Considerations

The development of any AI-based or identity-related application demands a rigorously ethical approach. Persona has publicly documented and rigorously adheres to a set of AI principles for the ethical development and conscientious deployment of its technology.

Ethical Development: Training Data

Persona Age Estimation has been developed using over one million ethically sourced images. All images used for training and benchmarking were collected with appropriate consent by Persona and its technology partner. This includes purpose-built data collection, where photographers were commissioned to capture individuals across varied environmental conditions.

No Persona customer data is used in either model training or benchmarking.

Ethical Development: Demographic Performance

Demographic Benchmarking and Fairness Evaluation

Persona has conducted extensive benchmarking of the Age Estimation model across multiple demographic dimensions, including self-reported age, gender, and race. This report provides a detailed breakdown of performance across these groups, demonstrating strong consistency and robustness across the demographic spectrum.

The core age estimation model—together with Persona-developed capabilities for secure capture, face quality assessment, and end-to-end verification—has undergone rigorous evaluation by both internal and external parties. External evaluations include participation in government programs such as the Australian Age Assurance Technology Trials, as well as testing conducted by independent laboratories, including NIST.

Benchmarking Methodology

For internal benchmarking, Persona conducts analyses informed by methodologies established by:

• NIST Face Recognition Vendor Tests (FRVT / FRTE)
• U.S. Department of Homeland Security (DHS) Science & Technology Directorate demographic evaluations

These institutions employ different approaches to demographic classification and fairness measurement. NIST primarily utilizes nationality-based population metadata, while DHS S&T combines self-reported race with color-calibrated photometers, an approach suited to controlled, in-person testing environments.

Persona has elected to use self-reported race rather than inferred skin tone or nationality. This approach aligns with established fairness best practices and reduces the risk of introducing unintended bias due to environmental or technical factors.

Demographic Group Definitions

For the purposes of internal benchmarking, race is categorized into four primary groups:

• Black
• South Asian
• East Asian
• White

Definitions of Black and White align with conventions established by the U.S. Census Bureau. The Asian category is subdivided into:

• South Asian, representing populations from the Indian subcontinent
• East Asian, representing other Asian populations

Ongoing Fairness Commitments

Persona remains committed to continuous improvement in fairness evaluation and the mitigation of differential performance, commonly referred to as algorithmic bias. As evaluation methodologies evolve, we will continue to incorporate leading frameworks and emerging research to refine demographic analyses and ensure equitable model performance across global populations.

Ethical Deployment: Privacy

Face images used for age estimation are classified as Personally Identifiable Information (PII) and are handled with a high degree of care.

The age estimation system does not learn from operational data. Images submitted during normal operation, and the resulting age estimates, are never used for model training or fine-tuning within Persona’s production environment.

Model improvements are identified through offline collaboration between Persona and its technology partner. Performance gaps are addressed by updating model architectures or training datasets using properly consented, carefully curated datasets that are fully independent of Persona’s customer traffic and services.

Updated models are trained offline and deployed through standard software release processes. This approach ensures strong separation between production data and model development, supporting privacy, data minimisation, and regulatory compliance.

Ethical Deployment: Use Case Analysis

Persona applies a consultative approach when deploying age estimation, evaluating whether proposed use cases align with ethical and privacy standards. Applications that do not meet these standards may be declined.

---

How Persona Age Estimation Works

1. Face Image (Selfie) Capture

Age estimation is performed on a selfie image, typically captured on a smartphone. Automated image quality metrics assess factors such as lighting, blur, and face size. If conditions are not optimal, the system provides real-time guidance (e.g., adjust lighting or camera position). Once quality thresholds are met, the image is captured and transmitted securely for analysis.

2. Age Estimation

The captured image is processed by Persona’s cloud-based age estimation software. The AI model, based on a Convolutional Neural Network, outputs an estimated age. This estimate can be combined with other identity signals depending on the configured workflow.

3. Policy Orchestration

Persona’s orchestration engine applies customer-defined policies to the age estimate. Policies may include age thresholds, escalation paths to alternative verification methods, and rules that vary by geography or use case.

Accuracy Analysis

Evaluation Sets

Persona evaluated its age estimation models using multiple datasets:

• Evaluation Set 1: Third-party benchmark dataset (N ≈ 200,000)
• Evaluation Set 2: UK Age Estimation Conformity Assessment (N ≈ 2,000)
• Evaluation Set 3: Australian Age Assurance Technology Trials (N ≈ 2,800)

Mean Absolute Error (MAE)

Source: Evaluation set 1
3rd-party partner benchmark dataset (N = 200,000)

Age range	Mean Absolute Error (MAE)
6–12	1.29
13–17	1.77
18–24	2.30
25–70	3.63
6–70	2.25

---

Source: Evaluation set 2
United Kingdom Age Estimation Conformity Assessment (N = 2,000)

Age group	Mean Absolute Error (MAE)	Standard Deviation of Absolute Error (SD)
<18	1.41	1.04
18–20	1.76	1.43

---

Source: Evaluation set 3
Australian Age Assurance Technology Trials (N ≈ 2,800)

Age range	Mean Absolute Error (MAE)
10–12	1.47
13–17	1.84
18–20	1.51

Demographic Analysis

MAE was evaluated across gender and race categories (Black, White, South Asian, East Asian). Results show relatively small variance across demographic groups within each age range, indicating consistent performance.

Mean Absolute Error (MAE) with Demographic Analysis

Source: Evaluation set 1
3rd-party partner benchmark dataset (N = 200,000)

Female — Mean Absolute Error (MAE)

Age range	Black	South Asian	East Asian	White	Overall
6–12	1.33	1.41	1.08	1.56	1.34
13–17	1.74	1.67	1.89	1.59	1.72
18–24	1.90	2.72	2.44	2.46	2.38
25–70	4.15	3.49	4.19	3.55	3.84
0–70 (Overall)	2.28	2.32	2.40	2.29	2.32

Male — Mean Absolute Error (MAE)

Age range	Black	South Asian	East Asian	White	Overall
6–12	1.34	1.25	1.25	1.21	1.23
13–17	1.72	1.81	1.91	1.63	1.81
18–24	2.38	2.86	2.06	2.04	2.23
25–70	3.84	3.21	3.69	3.23	3.42
0–70 (Overall)	2.32	2.29	2.20	2.03	2.17

Overall MAE

Age range	MAE
6–12	1.29
13–17	1.77
18–24	2.30
25–70	3.63
0–70 (Overall)	2.25

Here, the “Overall” values represent the average MAE across each of the various demographics, whether they be age, gender, or race (or some combination thereof). This table is represented visually in the following graph

Skin Tone Analysis

Mean Absolute Error (MAE) with Demographic Analysis

Source: Evaluation set 3
Australian Age Assurance Technology Trials

---

Methodology

Bias was assessed using parity-based analysis across multiple metrics for demographic subgroups defined by skin tone. Skin tones were classified according to the Fitzpatrick skin type scale, ranging from Type I (pale white) to Type VI (very dark brown/black).

Parity metrics quantify disparity in Mean Absolute Error (MAE) by comparing the performance of each skin tone group against the average MAE across all skin types.

Bias severity thresholds were defined as follows:

• Low: < 0.25
• Medium: 0.25 – 0.75
• High: > 0.75

---

Results

Persona was evaluated to have strong performance with low to moderate bias for darker skin tones. In some cases, observed bias favored darker skin tones, meaning the system achieved higher accuracy for darker skin tones compared to lighter skin tones—particularly for 18+ age gates and overall false negative rates.

Skin Type (Fitzpatrick)	Samples	MAE	MAE Parity	MAE Bias
I–II	989	1.39	0.11	Low
III–IV	1,027	1.52	0.02	Low
V–VI	206	1.95	0.45	Medium

---

Rankings

Although the Australian Age Assurance Technology Trial was not designed to formally rank providers, Persona’s solution demonstrated industry-leading performance in bias mitigation, achieving one of the two lowest MAE scores across demographic groups and evaluated use cases.

Persona achieved the lowest MAE for users under the age of 13. The only notable anomaly was a single MAE value of 2.5 for the 15-year-old age group, driven by one outlier in the dataset. The trial dataset intentionally included a higher representation of individuals from Indigenous communities, and one 15-year-old Indigenous participant was incorrectly estimated to be 62 years old, disproportionately influencing the average.

Under typical real-world conditions, MAE would be expected to be lower, as the Australian Age Trial intentionally oversampled minority populations to support robust bias evaluation.

False Positive Rate (FPR)

FPR varies based on the age threshold applied:

• At a threshold of 13, approximately 2.2% of users aged 6–11 were incorrectly classified as 13 or older.
• At a threshold of 25 (Challenge 25), 0% of users under 18 and ~2% of users aged 18–20 were estimated as 25 or older.
• Results support the use of configurable buffers between estimated age and enforced thresholds.

Comparative analysis from the Australian Age Assurance Technology Trials shows Persona performing competitively, particularly for younger age groups.

Generally Persona recommends a buffer age threshold of 2.5-3 to meet high efficacy

---

Third-Party Evaluations

Persona’s age estimation and broader age assurance solutions have been evaluated by multiple independent organizations, including:

• UK Age Check Certification Scheme (ACCS – Level 3)
• Australian Age Assurance Technology Trial
• Kantara Initiative
• National Institute of Standards and Technology (NIST)
• iBeta (ISO/IEC 30107-3)

These evaluations assess accuracy, bias, and compliance with relevant standards.

---

Summary

Persona Selfie Age Estimation combines AI-based age estimation with ethical development practices, privacy protections, and policy-driven orchestration. The solution is designed to support scalable age assurance across diverse use cases while maintaining consistent performance across demographics and complying with global regulatory expectations.

---

© Persona Identities, Inc.
This document is provided for informational purposes only and does not constitute legal advice.