Idiap researchers delve into biometric bias, template protection and PAD for cars

Biometrics experts at the Idiap Research Institute in Switzerland have contributed three papers to the latest issue of the identity-focused journal from the Institute of Electrical and Electronics Engineers.

The articles, written by researchers from Idiap’s Biometrics Security & Privacy group, are three of ten published in the January 2022 issue of the journal ‘IEEE Transactions on Biometrics, Behavior, and Identity Science.’

‘Fairness in Biometrics: A Figure of Merit to Assess Biometric Verification Systems’ was authored by Thiago de Freitas Pereira and Sébastien Marcel, and introduces the metric ‘Fairness Discrepancy Rate.’ The potential of the rate, referred to as FDR, is considered with a demonstration using two synthetic biometric systems. The metric was then tested for evaluations of gender and race demographics using face biometrics and three public datasets.

The reason to search for a new metric, according to the researchers, is that the existing methods, based on DET or ROC curves, assume demographic-specific decision thresholds, which Pereira and Marcel write are “not feasible or ethical in operational conditions.”

“We could observe via the FDR plots that all evaluated face verification systems presents (sic) gender and racial biases to some degree,” they conclude. “Furthermore, it was possible to quickly compare different face recognition systems concerning their demographic discrepancies using the Area Under FDR.”

FDR and Area Under FDR do not function as direct proxies for biometric verification accuracy, the authors note.

Sébastien Marcel also co-authored the other two papers, ‘Towards Protecting Face Embeddings in Mobile Face Verification Scenarios’ with Vedrana Krivokuca and ‘Domain-Specific Adaptation of CNN for Detecting Face Presentation Attacks in NIR’ with Ketan Kotwal, Sushil Bhattacharjee, Philip Abbet, Zohreh Mostaani, Huang Wei, Xu Wenkang and Zhao Yaxi.

The paper on protecting mobile face verification describes the production of more secure face biometric templates with “mapping based on multivariate polynomials parameterised by user-specific coefficients and exponents.” The researchers call the method PolyProtect, and say it can be tuned to an appropriate balance between recognition accuracy and the irreversibility of the templates.

The research into NIR in facial presentation attacks extends previous research on PAD systems for automobiles, considering a “lightweight face PAD framework” with “a 9-layer convolutional neural network (CNN).” The system developed returned an overall accuracy rate of 98 percent with the customized dataset, called VFPAD, which will be shared with the research community.

Related topics being tackled by Idiap researchers include synthetic biometric training data and biohashing.

Article Topics

bias  |  biometrics  |  face biometrics  |  Idiap  |  IEEE  |  presentation attack detection  |  synthetic data