02 Jul 2024

Buyers Guide

Complete playbook to understand liveness detection industry

Learn More

What is BPCER (Bonafide Presentation Classification Error Rate)?

Author: teresa_myers | 02 Jul 2024

In the previous knowledge base, we expounded APCER (Attack Presentation Classification Error Rate). We illustrated the findings of NIST tests in which we summarized the results and concluded that APCER can be calculated once BPCER (Bona Fide Presentation Classification Error Rate) is calculated, fixed and its standard value is set.

Let us describe BPCER in detail as we proceed with our astounding knowledge base series for biometric IDV enthusiasts.

Defining BPCER

BPCER stands for Bona Fide Presentation Classification Error Rate, an error rate of non-morphed (genuine) images falsely accepted as morphs. It is also called the ‘False Detection rate’. BPCER Error Rates occur when a system fails to distinguish between a genuine and a morphed image and will flag genuine user facial images as a facial morph attempt.

Calculating BPCER

When a biometric identity verification system such as a facial recognition solution’s algorithms is presented with morphed or bona fide images, it can make 2 major types of error including APCER and BPCER. NIST (National Institute of Standards & Technology) states that,

“When an algorithm fails to process an image as a bona fide one, it treats the morph image detection with a confidence score of 1 which is then used to calculate BPCER which is also a standalone quantity reported by NIST.”

Here’s the general formula for calculating BPCER:

BPCER (Bonafide Presentation Classification Error Rate) is calculated as the number of bonafide images ‘B’ divided by the total number of presented images ‘Nb’.

BPCER vs. APCER

BPCER is calculated independently of APCER. However, in a report on PAD (Presentation Attack Detection) including both stills and videos, NIST reported both APCER and BPCER each by fixing the error rate of the other at 0.01 i.e. APCER @ BPCER = 0.01 and BPCER @ APCER = 0.01.

Look at the BPCER threshold calculation below:

BPCER Formula

It is important to note that both BPCER and APCER have an impact on the accuracy and robustness of a face recognition and identity spoofing detection system.

Minimizing BPCER

To minimize BPCER, the Face Analysis Technology Evaluation (FATE) Part 10: Performance of Passive, Software-Based Presentation Attack Detection (PAD) Algorithms report aims to explore the effects of a technique called fusion which is an attempt to combine information from multiple sources to improve the accuracy levels of a biometric system by reducing the BPCER and APCER. NIST used a simple sum rule to summarize multiple algorithms’ BPCER and APCER rates.

Both Tables 26 and 27 show summarized results of 4 algorithms (each) for fusion.

BPCER @ APCER = 0.01 for fusion showed:

  • Maximum value = 0.193
  • Minimum value = 0.000
  • Average = 0.037

Demographic Considerations

The report also calculated BPCER for racial and color differences appearing in facial images of different individuals. BPCER’s value for white male photos was fixed at 0.03 due to large data sets (nearly 6000 images presented to the algorithms). However, NIST expects lower numbers of false detection BPCER in presentation attack detection rates.

Apart from this, the BPCER calculation for Face Morphing is also calculated by NIST and published. 

Can a Biometric Solution achieve 0/0 in BPCER vs. APCER?

So far, it is impossible for any biometric solution to achieve a perfect 0/0 score in BPCER vs. APCER calculation. This is because a minimum threshold is set for both at 0.01 and a trade-off is always there. It means that if the BPCER result for a particular algorithm comes ‘0’ the APCER would have a positive value and vice versa. This is the trade-off that occurs due to multiple reasons such as:

  • Cost of enhancing algorithm’s performance
  • Machine error
  • Anomalies and faults in datasets
  • Limitations in algorithms

Impact of High BPCER Levels

Reducing BPCER levels is critically important to ensure the algorithm’s accuracy. If BPCER levels start rising, it means that the PAs (Presentation attacks) have become more sophisticated like they did when the Gen-AI Deepfakes were introduced as a type of PA. Now AI deepfake attacks are the most threatening type of PAs for which the algorithms must be advanced.

BPCER or APCER levels will automatically rise when deepfake attacks evolve to a new level or a new form of presentation. This will not only decrease the usability of the biometric system but also raise questions about the identity verification service vendor as a whole. 

Following the NIST guidelines, trying fusion, maintaining the pace for PAD, and maintaining the best trade-off level between BPCER vs. APCER requires efforts including improving the Liveness Detection feature.