Architectural Bias in Face Presentation Attack Detection: A Comparative Study of Vision Transformers and Convolutional Neural Networks

2026-06-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

A comparative empirical investigation into Face Presentation Attack Detection (PAD) systems reveals that Vision Transformer architectures significantly reduce demographic bias compared to convolutional neural networks. Experiments on the CASIA-SURF Cross-Ethnicity Face Anti-Spoofing (CeFA) dataset evaluated a Multimodal ViT-Tiny, a ResNet18 CNN baseline, and a pretrained DeiT-S. DeiT-S achieved the highest overall accuracy of 97.27% and the lowest Equal Error Rate (EER) of 0.86%, surpassing ResNet18's 90.15% accuracy. Notably, DeiT-S reduced the inter-ethnic ACER gap between African and East Asian subjects to 0.13%, an 83% reduction from a reported 0.75%. Furthermore, DeiT-S demonstrated a 3.6x generalization advantage on zero-shot Central Asian subjects, maintaining 2.89% BPCER compared to ResNet18's 10.44%. These findings suggest that architectural design, particularly pretrained Vision Transformers, influences cross-demographic fairness in PAD systems.

Key takeaway

For AI Security Engineers designing or deploying face Presentation Attack Detection systems, you should prioritize pretrained Vision Transformer architectures. These models, like DeiT-S, demonstrate superior accuracy (97.27%) and significantly reduce demographic performance disparities, achieving an 83% reduction in inter-ethnic ACER gaps. Your systems will also benefit from 3.6x better generalization to unseen demographic groups, enhancing overall security and equity in biometric authentication.

Key insights

Pretrained Vision Transformers significantly reduce demographic bias and improve generalization in face Presentation Attack Detection.

Principles

• Pretrained Vision Transformers achieve superior PAD accuracy.
• They produce smaller demographic performance gaps.
• They generalize more equitably across unseen groups.

Method

Conducted a comparative empirical investigation of Multimodal ViT-Tiny, ResNet18, and pretrained DeiT-S architectures on the CASIA-SURF CeFA dataset.

In practice

• Consider pretrained Vision Transformers for PAD systems.
• Evaluate PAD systems for inter-ethnic ACER gaps.

Topics

• Face Presentation Attack Detection
• Vision Transformers
• Demographic Bias
• Biometric Security
• DeiT-S
• ResNet18

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, AI Security Engineer

Related on AIssential

• Find the Differences: Differential Morphing Attack Detection vs Face Recognition — Face recognition and differential morphing attack detection are functionally similar, and FR systems can be adapted for morphing detection.
• Adversarial Flow Matching for Imperceptible Attacks on End-to-End Autonomous Driving — AFM leverages Flow Matching to create imperceptible, transferable gray-box attacks against Transformer-based autonomous driving systems.
• From Vision to Text: A Compact Multimodal Approach for Robust, Cross-Domain Presentation Attack Detection on ID Cards — Compact multimodal PAD models need real-world data and fine-tuning for robust cross-domain performance, challenging synthetic dataset utility.
• Powerful AI is making facial recognition better at identifying you — Deep learning has made facial recognition over 99% accurate in controlled settings, despite ongoing bias and environmental challenges.
• Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery — Proto-LeakNet exploits diffusion model "signal leaks" in latent space for robust, interpretable synthetic image source attribution.
• Small Model, Big Results: How a 5MB On-Device NSFW Detector Outperforms Cloud APIs and Passes a… — Small, on-device object detection models can surpass large cloud APIs in NSFW accuracy and fairness by focusing on anatomical shapes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.