Editing Facial recognition system (section)

=== Cross-race effect bias ===

Facial recognition systems often demonstrate lower accuracy when identifying individuals with non-Eurocentric facial features. Known as the [[Cross-race effect]], this bias occurs when systems perform better on racial or ethnic groups that are overrepresented in their training data, resulting in reduced accuracy for underrepresented groups.<ref name="CrossRace2023">{{Cite arXiv |eprint=2305.16443 |title=Human-Machine Comparison for Cross-Race Face Verification: Race Bias at the Upper Limits of Performance |last1=Jeckeln |first1=Gabriel |year=2023|class=cs.CV }}</ref> The overrepresented group is generally the more populous group in the location that the model is being developed. For example, models developed in Asian cultures generally perform better on Asian facial features than Eurocentric facial features due to overrepresentation in the developers training dataset. The opposite is observed in models developed in Eurocentric cultures.<ref>{{Cite journal |last1=Phillips |first1=P. Jonathon |last2=Jiang |first2=Fang |last3=Narvekar |first3=Abhijit |last4=Ayyad |first4=Julianne |last5=O'Toole |first5=Alice J. |date=2011-02-02 |title=An other-race effect for face recognition algorithms |url=https://dl.acm.org/doi/10.1145/1870076.1870082 |journal=ACM Trans. Appl. Percept. |volume=8 |issue=2 |pages=14:1–14:11 |doi=10.1145/1870076.1870082 |issn=1544-3558|url-access=subscription }}</ref>

The systems used for facial recognition often lack the sufficient training needed to fully recognize those features not of Eurocentric descent. When the training and databases for these [[Machine learning|Machine Learning]] (ML) models do not contain a diverse representation, the models fail to identify the missed population, adding to their racial biases.<ref name=":1" /> 

The cross-race effect is not exclusive to machines; humans also experience difficulty recognizing faces from racial or ethnic groups different from their own. This is an example of inherent human biases being perpetuated in training datasets.<ref name="Sangrigoli2004">{{Cite journal |last1=Sangrigoli |first1=Sophie |last2=de Schonen |first2=Scania |title=Recognition of Own-Race and Other-Race Faces by Three-Month-Old Infants |journal=Journal of Child Psychology and Psychiatry |volume=45 |issue=7 |year=2004 |pages=1219–1227 |doi=10.1111/j.1469-7610.2004.00319.x|pmid=15335342 }}</ref>