Evaluation Metrics
For the performance evaluation, we selected the recently standardized ISO/IEC 30107-3 metrics: Attack Presentation Classification Error Rate (APCER), Normal/Bona Fide Presentation Classification Error Rate (NPCER/BPCER) and Average Classification Error Rate (ACER) as the evaluation metric, in which APCER and BPCER/NPCER are used to measure the error rate of fake or live samples, respectively. Inspired by face recognition, the Receiver Operating Characteristic (ROC) curve is introduced for large-scale face Anti-spoofing in our dataset, which can be used to select a suitable threshold to trade off the False Acceptance Rate (FAR) and False Rejection Rate (FRR) according to the requirement of real applications.
We will also give other metrics, such as 1-FRR (meaning correct acceptance rate) @FAR=10E-2, 10E-3, and ACER. About the definition of ROC, FAR, and FRR, please refer to [1].
Other metrics used for face Anti-spoofing will also be given:
(Positive samples mean attack ones, and negative samples mean Real/Bona Fide samples.)
Attack Presentation Classification Error Rate (APCER ):
APCER = FN / (TP + FN)
Normal (Bona Fide) Presentation Classification Error Rate (NPCER/BPCER ):
NPCER/BPCER= FP/(FP + TN)
Average Classification Error Rate (ACER):
ACER = (APCER + NPCER) / 2
False Rejection Rate (FRR):
FRR = FN/(TP+FN)
False Acceptance Rate (FAR):
FAR = FP/(FP+TN)
Note:
TP: The attacks are recognized as the attacks;
TN: The real samples are recognized as the real ones;
FP: The real samples are recognized as the attacks;
FN: The attacks are recognized as the real samples;
ref:
[1] Aghajan, H., Augusto, J. C., & Delgado, R. L. C. (Eds.). (2009). Human-centric interfaces for ambient intelligence. Academic Press. [link]