Face Recognition Algorithm Bias: Performance Differences on Images of Children and Adults

Nisha Srinivas, Karl Ricanek, Dana Michalski, David S. Bolme, Michael KingWorkshops @ CVPR 2019 · DOI

Abstract

Current state-of-the-art face recognition (FR) systems are predominantly trained on adult datasets, leading to an implicit assumption that these models will generalize to all age groups. This study benchmarks eight top-tier FR algorithms—including commercial, government, and open-source solutions—to quantify the performance gap when processing images of children. Our results reveal a universal "negative bias" against children, driven by non-linear biological growth and a lack of representative training data. We demonstrate that while individual systems struggle, score-level fusion techniques can significantly mitigate this performance decline.

Introduction: The Challenge of Child Face Recognition

Most face recognition research operates under the premise that a model optimized for adults will naturally extend to children. However, children are not merely "scaled-down" versions of adults. Their craniofacial growth follows complex biological mechanisms that result in non-linear changes to facial proportions and features over time.

Figure 1: Examples of age progression in children from the ITWCC dataset, illustrating the significant facial changes over short periods.

This biological reality translates into a significant algorithmic bias. In high-stakes scenarios—such as identifying victims of child exploitation, locating abducted children, or managing border security—the failure of an algorithm to accurately recognize a child's face has severe real-world consequences. This paper seeks to quantify this performance gap across a diverse range of modern FR systems.

Methodology and Datasets

The researchers conducted a rigorous benchmarking study using eight FR systems: five commercial-off-the-shelf (COTS) systems, two government-off-the-shelf (GOTS) systems, and one open-source solution based on VGG-Face. Evaluation was performed across two primary protocols:

  1. Verification (1:1 matching): Confirming if two images represent the same identity.
  2. Identification (1:N matching): Searching a database to find a matching identity.

To ensure real-world relevance, the study utilized two "in-the-wild" datasets characterized by unconstrained poses, lighting, and expressions:

  • ITWCC-D1 (In The Wild Child Celebrity): A refined subset containing 7,990 images of 745 child subjects (celebrities and actors) across various ages.
  • LFW-D1 (Labeled Faces in the Wild): A subset of the industry-standard LFW dataset, serving as the adult baseline with 9,164 images of 1,680 subjects.

Benchmarking Results

Universal Negative Bias

Every algorithm tested performed significantly worse on children than on adults. This confirms that the "age gap" in FR performance is a systemic industry-wide issue rather than a flaw of specific vendors. The primary cause is the overlap in score distributions; while adult genuine and imposter scores are well-separated, those for children overlap significantly, making it difficult for systems to distinguish between different children or confirm the identity of the same child over time.

Figure 2: ROC curves for eight face recognition algorithms (FR-A through FR-H) on the ITWCC-D1 child dataset, showing varied performance levels.

Performance Metrics

As shown in the ROC analysis, performance varied widely among the systems. For child verification, FR-E emerged as the most accurate. In identification tasks, FR-G led the group with an 82% retrieval rate—outperforming the weakest system in the study by nearly 28%. A notable "inflection point" was observed around age 8; performance is particularly poor for younger cohorts and begins to stabilize as facial structures become more adult-like.

Mitigation via Score-Level Fusion

To combat the performance decline, the authors tested six different score-level fusion schemes (labeled Fusion A-F). These schemes combined the outputs of multiple algorithms using various normalization techniques (z-score vs. min-max) and fusion rules (sum, max, or min).

Fusion-B (min-max normalization with a sum rule) was the most effective strategy. It improved the True Accept Rate (at 0.1% FAR) by 2% and boosted Rank-1 identification accuracy by 4-5% compared to the best individual system. This suggests that different algorithms capture complementary child-specific features, making their collective intelligence more robust than any single "expert" system.

Conclusion and Future Work

The inherent bias in current FR systems stems largely from training data scarcity. Most top-tier models are heavily weighted toward adult features. Future improvements will require:

  • Diverse Training Data: Developing massive, diverse datasets specifically for child demographics.
  • Selective Fusion: Moving beyond "fuse-all" approaches to identify specific algorithm pairs that offer the best complementary results for children.
  • Targeted Retraining: Retraining open-source models on child data to eliminate bias at the source rather than relying solely on post-processing mitigation.

Summary updated

BibTeX
@inproceedings{srinivas2019face,
  author = {Nisha Srinivas and Karl Ricanek and Dana Michalski and David S. Bolme and Michael King},
  title = {{Face Recognition Algorithm Bias: Performance Differences on Images of Children and Adults}},
  booktitle = {Workshops @ CVPR},
  year = {2019},
  doi = {10.1109/cvprw.2019.00280},
}
Figure