Trueface explains reducing bias in face biometrics model
Trueface says that the Fairface face biometrics dataset has helped quantify bias factors in its TFV5 facial recognition model, enabling the company to reduce differences in accuracy between ethnicities. Cyrus Behroozi, a computer vision software developer at the company, elaborated in a post on Medium the results of Trueface’s evaluation of TFV5.
The post also provides a blueprint that guides clients to reduce false positives in their facial recognition applications by raising their models’ operational thresholds.
In the post, Behroozi lays out the quantifiable differences between the TFV5 model and its predecessor the TFV4. One significant difference is a decrease in bias across all ethnic and gender groups, including East Asians and Southeast Asians, which Behroozi says are typically underrepresented. This is due to the addition of supplementary images from underrepresented groups to the training dataset.
“The Fairface dataset contains a balanced number of face images from seven major ethnic groups and contains no more than a single image for each identity. In the evaluation, we generate a face recognition template for each image in the dataset, then compare every face template against one another to generate a similarity score,” he explained.
When compared to the TFV4, the new model appears to yield a significant decrease in bias in historically underrepresented ethnic groups from East and South Asia. Behroozi notes that this is mainly due to the ethically sourced biometric training dataset that was comprised of images of these groups.
Behroozi also noted that this can bring technological equity regardless of gender and ethnicity. He further adds that such a reduction in false positives could also reduce security risks when applied to biometric access control scenarios. “In general, we advise that our clients operate at a similarity score threshold of between 0.3 to 0.4, though the exact threshold is ultimately dictated by the desired False Positive Rate or False Negative Rate. What you will notice in the two plots below is that TFV5 has significantly fewer false positives in the operating region for all ethnicities,” he added.
This post was a follow-up to Trueface’s initial review of the bias factor within its biometric facial recognition model published in 2020. In it, Behroozi laid out how the evaluation with Fairface was designed.
This post was updated at 6:33pm on Tuesday, March 30, 2020 to clarify the details of the changes made to the algorithm.