Online Age Verification Software

AI Agents. Age Validation Solution (KYC)

The ML model, trained on open-source images, identifies ages with 85% accuracy and processes up to 350,000 users each day

Project Highlights

AI-Powered Age Validation: Developed a machine learning solution for accurate user age validation using convolutional neural networks and variational autoencoders.

Open-Source Training Data: Built and trained the model exclusively with publicly available data due to security restrictions on real user data.

Feature Extraction: Extracted facial landmarks like eye distance and forehead height, improving predictions.

Obstruction Identification: Added flags to detect obstructive elements like glasses, hats, makeup, and cigarettes.

High Accuracy: Achieved 80-85% accuracy across various age groups.

Large-Scale Processing: Processed 250,000-350,000 users daily.

Outperforming Mastercard’s Ekata 4x times.

Industry: Social media app

Location: Global

Tech Stack:

Convolutional Neural Networks
Autoencoders
Variational Autoencoder
TensorFlow
SciPy
OpenCV
Computer Vision

Offered Services: AI/ML Development & Solution Architecture

Collaboration Model: Time & Material

Platforms: Web

Product Overview

The Challenge

‍

The business needed an AI solution capable of verifying the ages of over 300 million users with minimal human interaction. However, access to real user data for training the machine learning model was restricted due to privacy and security concerns. This required the team to develop and train their model using publicly available datasets, which posed a unique challenge. We had to accurately tag age data and balance different age groups across classes. Moreover, obstructive elements like glasses, hats, and makeup made facial feature identification more difficult. The solution demanded a precise approach to age identification, even with limited and imbalanced data.

The Core Solution

We approached the challenge with innovative precision and bold confidence, crafting a multi-layered ML solution that left no stone unturned.

1. Data Collection & Preparation

Given the security concerns around user data, we curated a robust training set using publicly available photos, meticulously annotated with accurate age markers.

Our team tackled the shortage of proprietary data with creativity and diligence, balancing the dataset to reflect realistic age distributions.

2. Initial Age Categorization

To establish a foundation, we crafted a cascade model with broad age group distinctions:

Children (<12 years)
‍Distinct juvenile facial traits made this group relatively straightforward to identify.

Youth & Adults (12-40 years)
‍This segment presented nuanced adolescent characteristics, requiring a more nuanced approach.

Seniors (>40 years)
‍Strong facial markers typical of aging facilitated accurate categorization.

With this high-level categorization, we ensured the model was finely attuned to the major differences in these broad age ranges.

3. Detailed Age Segmentation

Recognizing that the 12-40 age group was the most challenging to classify, we created a second layer that segmented this category further.

12-15 years: Early adolescent features.

15-16 years: Transitioning adolescence.

16-20 years: Older teens and young adults.

25+ years: Full-fledged adults.

The sub-models filtered their data, enhancing precision and enabling the system to focus on the most distinguishing characteristics for each sub-group.

4. Age-Specific & Legal Boundary Classification

For ultimate granularity, two specialized models were employed:

Age-Specific Classification
‍Isolated ages between 15-20 years with remarkable clarity, overcoming subtle differences.

Legal Split Classification
‍Categorized users with strong confidence as being either below or above 18 years.

5. Facial Landmark Analysis

Our final layer went a step further by mapping crucial facial landmarks like eye distance and forehead height while deftly filtering obstructive elements like glasses and hats.

By implementing this sophisticated approach, we achieved impressive accuracy in age identification across a wide range of users and the ability to process a high volume of requests with minimal human intervention.

The outcome

Our advanced AI solution achieved remarkable results, reaching a consistent 80%-85% accuracy in age verification across various demographic groups. The model efficiently processed up to 350,000 user images daily, categorizing them based on predefined age ranges and significantly minimizing errors. This impressive achievement left the client delighted and confident in the system's capabilities.

We aim to elevate this solution further by training the model on real user data, enabling us to enhance the model’s predictive performance and reach around 95%+ accuracy level. This comprehensive approach will provide more reliable verification, reducing edge cases and creating a seamless, automated process that scales effortlessly.

In addition to improving accuracy, the next step is to integrate a sophisticated classification system for explicit content. By implementing this feature, the model will categorize images across different explicit content types, ensuring adherence to compliance standards and enhancing the safety of the platform. This level of classification will be a valuable addition, enabling nuanced detection of inappropriate content while fortifying the age verification process. Our team is dedicated to providing clients with high-performing, efficient, and secure AI solutions that can adapt to their evolving needs.