Facial Anti-Spoofing Using "Clue Maps"

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Facial recognition systems are vulnerable to spoofing attacks, posing risks to online financial security.
• Existing multi-modality anti-spoofing methods can be costly due to data acquisition requirements.
• There is a critical need for robust anti-spoofing solutions with strong generalization capabilities.

Purpose of the Study:

• To propose a novel representation learning method for facial anti-spoofing.
• To develop a cost-effective solution that overcomes limitations of multi-modality approaches.
• To enhance the generalization ability of spoofing attack detection models.

Main Methods:

• A representation learning method utilizing an Auto-Encoder structure based on Swin Transformer and ResNet.
• Supervised training with a combination of cross-entropy loss, semi-hard triplet loss, and Smooth L1 pixel-wise loss.
• An architecture comprising an Encoder for feature extraction, a Decoder for generating "Clue Maps", and an auxiliary classifier for decision-making.

Main Results:

• The proposed model achieved superior performance on popular spoofing databases (CelebA, OULU, CASIA-MFSD).
• Intra-dataset experiments yielded low Average Classification Error Rates (ACER) of 1.2% and 1.6%.
• Inter-dataset experiments set a new state-of-the-art performance with a 23.8% Half Total Error Rate (HTER) on the Replay-attack dataset.

Conclusions:

• The developed Auto-Encoder based representation learning method effectively detects facial spoofing attacks.
• The approach demonstrates superior generalization ability and outperforms existing anti-spoofing models.
• This research contributes a significant advancement in securing facial recognition systems against presentation attacks.