Method for facial recognition bias reduction with adversarial network shows promise

A prominent trio of biometrics researchers have proposed a way to remove the difference, or bias, from facial recognition performance between different demographics.

A paper jointly written by Sixue Gong, Xiaoming Liu and Anil K. Jain, all of Michigan State University, ‘Jointly de-biasing face recognition and demographic attribute estimation,’ was presented at the European Conference on Computer Vision (ECCV) 2020.

The researchers propose a novel de-biasing adversarial network (DebFace), which “learns to extract disentangled feature representations for both unbiased face recognition and demographics estimation.”

The network proposed in the paper is made up of one identity classifier and three demographic classifiers, one each for gender, age, and race. Correlation among feature factors is minimized through adversarial learning to reduce the influence of factors associated with bias, and the researchers also designed a scheme for combining demographics with identity features to improve the demographic balance of faces represented.

The question of whether bias could be trained out of facial recognition without reducing overall accuracy to the lowest common denominator was recently raised by ID4Africa Executive Director Dr. Joseph Atick during the organization’s livecast ‘Spotlight on Face Recognition Technology.’

Improving overall performance without including additional bias is non-trivial, Sixue told Biometric Update in an email.

“Since demographic attribute is discriminative to identities (different races can’t be the same subject), its removal will inevitably lead to a more challenging setting for FR,” she explains. “DebFace sacrifices the accuracy for cohorts with a large number of face samples while it improves the accuracy of cohorts with less images.”

The overall results of their experiments were encouraging, with reduction in bias and improved demographics estimation with performance comparable to state-of-the-art systems, according to the paper.

“One strategy to reduce bias while keep the demographic information is to raise the feature discriminability for under-represented cohorts by adding extra capacity to the corresponding feature extraction functions, but remaining the ones for features of well-represented cohorts,” says Sixue. “In this way, the general performance on all the cohorts can be increased, and meanwhile, the gap of accuracy can be decreased between under-represented cohorts and well-represented cohorts.”

Compared to the baseline method, DebFace reduced bias in both biometric identity verification and demographic estimation for gender, age and race. While overall accuracy was reduced, DebFace-ID had facial verification accuracy with the RFW (Racial Faces in-the-Wild) dataset ranging from 93.67 percent on Black faces to 95.95 percent on white faces in testing by Sixue, Xiaoming and Jain.

Sixue notes that balanced datasets have still been shown to produce bias, proving that other factors, such as camera settings, capture conditions, image qualities and demographic labels can also create bias. Even with demographic representation in mind from the start, it is difficult to build a truly balanced database because of what Sixue calls “the multiplicative effect of imbalance,” so other methods should also be applied to mitigate bias in facial recognition.

Where DebFace balances the trade-off of accuracy and bias by generating debiased representations for identity and demographics, blindly combining demographic features and identity could re-introduce bias. In future work, therefore, the researchers plan to experiment with an aggregation scheme combining demographics and identity without introducing bias through the dataset or algorithm.

“From the perspective of feature distribution, aggregating representations of race, gender, and identity is a process of re-mapping the identity feature representation back to its corresponding demographic area,” observes Sixue. “The identity features will be gathered into multiple clusters based on their demographic attributes (see Fig. 6 in the paper). The goal is to enhance the discrimination of feature points in each demographic group and narrow down the difference of feature distributions between demographic clusters.”

Work on reducing demographic differentials in face biometrics continues, with some of the brightest minds in the field applying themselves to the issue. The degree to which they are successful may go a long way towards determining how widespread facial recognition ultimately becomes.

Article Topics

accuracy  |  Anil Jain  |  biometric identification  |  biometric-bias  |  biometrics  |  biometrics research  |  dataset  |  identity verification  |  MSU  |  training