# Improving the Accuracy of Automated Facial Age Estimation to Aid CSEM Investigations

## Full Text

This Markdown is an automated text extraction for search and reading tools. The [hosted PDF](FacialAgeEstimationPoster.pdf) is the authoritative publication source.

Source PDF: [FacialAgeEstimationPoster.pdf](FacialAgeEstimationPoster.pdf)
Source SHA-256: `daee1f4c009800c84919682b31802901f96f7822ad1421338bb928ffca76b7e6`
Generated with `pdftotext -nopgbrk`; formatting and reading order may differ from the PDF.

---

Improving the Accuracy of Automated Facial
Age Estimation to Aid CSEM Investigation
Felix Anda, David Lillis, Aikaterini Kanta, Brett Becker, Elias Bou-Harb, Nhien An Le Khac, and Mark Scanlon
Forensics and Security Research Group, University College Dublin, Ireland.
felix.andabasabe@ucdconnect.ie, {mark.scanlon, an.lekhac}@ucd.ie

ABSTRACT

PRELIMINARY RESULTS

The investigation of child sexual exploitation material (CSEM) is one of the
more commonly encountered criminal investigation types throughout the
world. While hash lists of known illegal files are used to identify previously
encountered material on new devices, new previously unencountered
material requires expert, manual analysis and categorisation. The
discovery and analysis of these digital images and videos has the potential
to be greatly expedited with the use of automated artificial intelligence
(AI) based techniques. Intelligent, automated human-in-the-loop evidence
processing and evidence prioritisation could also help alleviate some of
the digital evidence backlogs that have become commonplace worldwide.
In order for AI aided CSEM investigations to be beneficial, the fundamental
question when analysing multimedia content becomes “how old is each
subject encountered?’’

While there are a number of methods that can contribute to easing the
digital evidence backlog [3][4], accuracy of automated age inference in
images and videos is one important technique needed to streamline CSEM
cases. This study evaluated four different age prediction services from
both online and offline sources. The four services evaluated were Amazon
Rekognition (AWS), Microsoft Azure, Deep Expectation (DEX), and HowOld.net. Initial evaluation results on an age range from 0 to 25, indicated
that AWS had the overall lowest error rate, followed by How-Old.net.
However, the ages that surround the borderline between minority and
adulthood (considered to be 18 for this study) were found to follow a
different pattern, where DEX surpassed the performance of AWS and
Azure. Following these observations, a separate dataset was curated
featuring a higher number of sample images in the age range of 1 to 25
inclusive.
Experiments on this dataset indicated that ensemble approaches based on
regression substantially outperformed the four systems used for this test.
Gradient Boosting and Bagging Regressor approaches not only
outperformed the best individual system for the borderline range (16-17)
by over 40 %, they were also superior to a hypothetical perfect system
that chooses the best prediction from those available each time.
Furthermore, the DS13K VGG-16 based model trained for this work had
the best figures for the borderline age range (16-17) and a competing third
place for the age range (0-5), as depicted in Figure 2. The overall
conclusion of the study is that even off-the-shelf regression techniques
have been demonstrated to improve upon the performance of commercial
offerings, by combining their outputs effectively. This offers a clear
motivation for further work on bringing AI based techniques to bear on
this and associated DF challenges.

Figure 1. Example of Borderline Adulthood Identification Challenge (Photo is CC0/Public Domain)

PROBLEM STATEMENT
Accurate facial age estimation has been an arduous task for both humans
and machines (Figure 1 depicts a sample challenge for both humans and
machines encountering borderline cases). Moreover, the influence of
factors such as environment, health habits, lifestyle, makeup, emotions,
uncontrolled lightning, and partial facial occlusion hinder the age
prediction process and thus impact the final result . Nevertheless, humans
are quite accurate at estimating the age of other humans. The mean
absolute error (MAE) rate has been measured to vary from 2.07 to 8.62
years depending on the age of the assessor and the studied subject, and
the difference between the two [1]. Conversely, machines have reached
MAE figures of approximately 4.1 [2] and vary according to the datasets
used in the implementation. The use of automation in DF has been
criticised due to the constant need of a human-in-the loop approach and
the lack of accuracy that models achieve for predicting age.

Figure 2. Age Estimator Performance per Age Group

Further information is contained
in our published paper on Age
Estimation entitled “Evaluating
Automated Facial Age Estimation
Techniques for Digital Forensics”:

METHODOLOGY
Our work presents the evaluation of existing cloud-based and offline age
estimation services, outlines our deep learning age estimation model,
DS13K (built with a VGG-16 Deep Convolutional Neural Network
architecture) and uses an ensemble technique to improve the accuracy of
underage subject age estimation. The contributing services consisted of
Amazon Rekognition, Microsoft Azure Cognitive Services, How-Old.net,
and Deep Expectation (DEX). It was found that for the borderline
adulthood age range of 16 to 17 years old, our deep learning model
performed best with a accuracy performance rate of 68%. A comparative
examination of the obtained results allowed us to identify performance
trends and issues inherent to each services/tool and develop ensemble
techniques to improve the accuracy of automated adulthood
determination.

REFERENCES
1.

2.

3.

4.

E. Moyse and S. Brédart. An own-age bias in age estimation of faces.
Revue Européenne de Psychologie Appliquée/European Review of Applied
Psychology, 62(1):3 – 7, 2012.
Khoa Luu, Keshav Seshadri, Marios Savvides, Tien D Bui, and Ching Y Suen.
Contourlet appearance model for facial age estimation. In Biometrics (IJCB),
2011 International Joint Conference on, pages 1–8. IEEE, 2011.
Anda, F., Lillis, D., Le-Khac, N-A., & Scanlon, M. (2018, May). Evaluating
Automated Facial Age Estimation Techniques for Digital Forensics. In 12th
International Workshop on Systematic Approaches to Digital Forensics
Engineering (SADFE), IEEE Security & Privacy Workshops. IEEE.
James, J. I., & Gladyshev, P. (2013). Challenges with Automation in Digital
Forensic Investigations. Retrieved from http://arxiv.org/abs/1303.4498