Onfido and Trueface work to remove bias from facial recognition technologies

Onfido and Trueface held two separate webinars at the International Face Performance Conference (IFPC) on Thursday, focusing on reducing bias from facial recognition applications.

At the conference, Martins Bruveris, machine learning research scientist at Onfido talked about reducing geographic performance differentials for facial recognition.

Mosalam Ebrahimi from Trueface, on the other hand, focused on discussing bias mitigation strategy in relation to overly confident false matches.

Reducing geographic performance differentials in face biometrics

Bruveris’ presentation opened by describing the general issues related to achieving 1:1 facial recognition rates between selfies and document photos.

The scientist mentioned the fact that photos in documents are not in high resolution, often are partially obscured by security features and holograms, and have low contrast.

Moreover, since they are printed on paper, they are subject to inherent image quality degradation over time.

Bruveris proceeded by saying that “no research happens in a vacuum,” and mentioning numerous sources upon which Onfido has run his tests in order to reduce geographic performance differentials in face biometrics.

To conduct these tests, the company also analysed an in-house dataset of 6.8 million image pairs, plus additional 100,000 external ones, including document issued IDs with available metadata regarding gender and other information.

Bruveris noted how the analysed data was ”imbalanced” and not uniformly distributed as the company had more datasets in countries where it had conducted business, mostly Europe and the U.S.

In other words, some countries were very well represented in the datasets, while others were not. This led to a false acceptance rate substantially higher in Asia and Africa.

In order to tackle this issue, Onfido deployed a series of mitigating strategies revolving around the ideas of Equal Sampling, Adjusted Sampling, and Dynamic Sampling.

Equal Sampling consists of choosing data samples in equal numbers from each continent, while weighted sampling attributes data sets with fewer entries a higher value.

Dynamic Sampling refers to weighted sampling with weights that are dynamically adjusted during training based on within-class false acceptance rate (FAR).

Bruveris specified that Onfido did not change the size of the dataset, but only the frequency with which a sample from each continent was chosen.

After the experiments, the company concluded that performance differentials can be reduced without balanced data, but that reducing FAR differential can lead to increased False Rejection Rate (FRR) differentials.

A facial recognition bias mitigation strategy for overly confident false matches

Mosalam Ebrahimi picked up where Bruveris left, reinforcing what the Onfido’s scientist said about algorithm bias in facial recognition applications.

He mentioned image recognition problems related to angle, grain and blurring. Moreover, that ethnicities underrepresented in the data set return higher false positives.

Ebrahimi gave a brief overview of Trueface’s work, mentioning some of the company’s recent projects. Specifically, the researcher described the company’s solution that generates a QR code containing encrypted biometric data extracted from a photo.

Ebrahimi also presented data related to the company’s algorithm speed, saying it is now four times faster than their own method in the FRVT 1:1 report on the same hardware.

This also allows the company’s software to selectively compress the parts of an image that have a higher bitrate using their own encoder and decoder, reducing file size, and maintaining high enough quality to reduce false positives.

The Trueface’s researcher then analysed why the similarity score (distance) is an imperfect confidence measure.

He argued the issue lies within the fact that the embedding is non-isometric with high distortion, and it maps the out-of-distribution points to random points in the destination metric.

Ebrahimi concluded his presentation by saying that overly confident false matches can be prevented by taking into consideration similarity scores’ probability density estimation, instead of point estimation.

Article Topics

accuracy  |  biometric bias  |  biometric identification  |  biometrics  |  biometrics research  |  facial recognition  |  NIST  |  Onfido  |  Trueface