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

Technical Summary: Architectural Bias in Face Presentation Attack Detection

Problem Statement
Face Presentation Attack Detection (PAD) systems are critical for securing biometric authentication but suffer from systematic demographic bias, disproportionately affecting individuals with darker skin tones. This bias is structural, stemming from underrepresentation in training datasets and the physical properties of image acquisition, where darker skin reflects significantly less light than lighter skin. While recent work using classical methods (e.g., Local Binary Patterns with ethnicity-aware preprocessing) has demonstrated that bias can be mitigated, it remains an open question whether architecturally advanced models, specifically Vision Transformers (ViTs), can achieve comparable or superior fairness. Unlike Convolutional Neural Networks (CNNs), which rely on local texture patterns and are known to be biased toward texture over global shape, ViTs utilize multi-head self-attention to capture global relationships. This study investigates whether the DeiT-S (Data-efficient Image Transformer Small) architecture inherently reduces demographic bias in PAD compared to CNN baselines and classical methods.

Methodology
The study conducts a comparative empirical investigation on the CASIA-SURF Cross-Ethnicity Face Anti-Spoofing (CeFA) dataset, which includes 1,607 subjects across African (AF), East Asian (EA), and Central Asian (CA) ethnic groups. The evaluation framework adheres to the statistical validation protocols established in prior work [6], employing McNemar's significance testing and bootstrap resampling.

Four models were evaluated:

1. LBP + SGD Baseline: A classical, fairness-aware model using multi-scale Local Binary Patterns with group-specific decision thresholds (results carried forward from [6]).
2. ResNet18 CNN: A standard CNN baseline pretrained on ImageNet and fine-tuned on CeFA.
3. Multimodal ViT-Tiny (Scratch): A transformer model trained from scratch on the CeFA dataset to test the inherent fairness of the architecture without pretraining.
4. DeiT-S: A pretrained Vision Transformer (ImageNet weights) fine-tuned on CeFA.

All models utilized identical ethnicity-aware preprocessing (CLAHE and adaptive gamma correction) to isolate architectural effects. The evaluation included standard PAD metrics (Accuracy, APCER, BPCER, ACER, EER, AUC) and fairness metrics, specifically the inter-ethnic accuracy gap and zero-shot generalization performance on the unseen Central Asian (CA) demographic group.

Key Results

• Overall Performance: DeiT-S achieved the highest overall accuracy (97.27%) and the lowest Equal Error Rate (0.86%), significantly outperforming the ResNet18 baseline (90.15% accuracy, 1.90% EER).
• Demographic Fairness (Seen Groups): DeiT-S reduced the inter-ethnic ACER gap between African and East Asian subjects to 0.13%. This represents an 83% reduction compared to the 0.75% gap reported in the LBP-based study [6]. In contrast, the ViT-Tiny model trained from scratch exhibited a 3.3× disparity in error rates between these groups, highlighting the instability of unpretrained transformers on small datasets.
• Zero-Shot Generalization: The most significant finding concerns generalization to unseen demographics. On the zero-shot Central Asian group, ResNet18 suffered a high Bona Fide Presentation Classification Error Rate (BPCER) of 10.44%, incorrectly rejecting genuine users at a high rate. DeiT-S maintained a BPCER of 2.89% on the same group, demonstrating a 3.6× generalization advantage.
• Accuracy vs. Fairness: The study found that high aggregate accuracy does not guarantee fairness. The scratch-trained ViT-Tiny achieved 96.48% accuracy but exhibited severe demographic disparities, whereas DeiT-S achieved both higher accuracy and superior fairness.

Significance and Claims
The paper argues that architectural design plays a measurable role in cross-demographic fairness in PAD systems. The results suggest that pretrained Vision Transformers, specifically DeiT-S, offer a dual advantage: superior detection accuracy and reduced demographic performance gaps. The authors attribute the improved fairness of DeiT-S to a combination of its global self-attention mechanism—which captures liveness cues transcending ethnicity-specific texture patterns—and the diversity of its large-scale ImageNet pretraining.

The study concludes that while CNNs may perform competitively on seen demographic groups, they fail to generalize equitably to unseen groups, creating "two-tier" access systems. In contrast, ViTs demonstrate a capacity for more equitable generalization. The authors modestly note that the observed fairness is likely a joint product of architecture and pretraining data diversity, and that fully isolating these factors requires further controlled ablation studies. The work does not claim to eliminate bias entirely but provides evidence that architectural choices can significantly mitigate it without relying solely on explicit fairness-aware preprocessing.