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

By Nisha Srinivas, Karl Ricanek Jr., Dana Michalski, David S. Bolme, Michael King. 2019

74 citations recorded by the source corpus.

In this work, we examine if current state-of-the-art deep learning face recognition systems exhibit a negative bias (i.e., poorer performance) for children when compared to the performance obtained on adults. The systems selected for this work are five top performing commercial-off-the-shelf face recognition systems, two government-off-the-shelf face recognition systems and one open-source face recognition solution. The datasets used to evaluate the performance of the systems are both unconstrained in age, pose, illumination, and expression and are publicly available. These datasets are indicative of photo journalistic face datasets published and evaluated on over the last few years. Our findings show a negative bias for each algorithm on children. Genuine and imposter distributions highlight the performance bias between the datasets further supporting the need for a deeper investigation into algorithm bias as a function of age cohorts. To combat the performance decline on the child demographic, several score-level fusion strategies were evaluated. This work identifies the best score-level fusion technique for the child demographic.

Subfields: Computer Vision and Pattern Recognition, AI Ethics, Fairness & Society, Signal Processing

Citations in the corpus

Saving Face: Investigating the Ethical Concerns of Facial Recognition Auditing (2020, neutral): “ng these datasets have heavily informed many of the recent attempts to address the dearth of diversity in FPT benchmarks and proposed frameworks for evaluating classification bias in deployed systems [32, 43, 45]. Although these prior studies evaluate commercially deployed models for gender and raci”
Face Recognition: Too Bias, or Not Too Bias? (2020, unknown): “d to train. In contrast, our BFW assumes VGG2 as the test set. Furthermore, BFW balance subgroups: RFW splits subgroups by gender and race, while BFW has gender, race, or both). Most similar to us is [8,18,22,33] - each was motivated by insufﬁcient paired data for studying bias in FR. Then, problems”