Vision Transformers slash facial recognition bias by 83% in new study
DeiT-S model achieves 97.27% accuracy with near-zero demographic fairness gap
A new study published on arXiv (2606.18510) by Ntung, Tuyisenge, and Ndibwile investigates how model architecture affects demographic bias in face presentation attack detection (PAD) – systems that spot fake faces trying to fool biometric scanners. The researchers tested three architectures on the CASIA-SURF Cross-Ethnicity Face Anti-Spoofing (CeFA) dataset: a ResNet18 CNN baseline, a small Vision Transformer (ViT-Tiny) trained from scratch, and a pretrained DeiT-S transformer fine-tuned on African, East Asian, and zero-shot Central Asian groups.
The results are striking. DeiT-S achieved 97.27% overall accuracy and a 0.86% equal error rate (EER), far surpassing ResNet18's 90.15% accuracy. But the real headline is fairness: DeiT-S reduced the inter-ethnic asymmetric error (ACER) gap between African and East Asian subjects to just 0.13%, compared to 0.75% in previous LBP-based work – a massive 83% reduction. Even more impressive, when tested on unseen Central Asian faces (zero-shot), ResNet18 had a 10.44% false positive rate while DeiT-S kept it to just 2.89%, a 3.6x generalization advantage. The authors argue that pretrained Vision Transformers inherently learn more equitable feature representations, suggesting architectural design is a key lever for reducing demographic bias in biometric systems.
- DeiT-S transformer achieved 97.27% accuracy and 0.86% EER, vs ResNet18's 90.15%
- Inter-ethnic fairness gap reduced to 0.13% ACER – an 83% improvement over prior methods
- 3.6x better zero-shot generalization: 2.89% BPCER on unseen Central Asian faces vs 10.44% for CNN
Why It Matters
Architecture choice can dramatically reduce demographic bias in facial anti-spoofing, making security systems fairer for all skin tones.