[Paper Review] Impact of Face Image Quality Estimation on Presentation Attack Detection

Carlos Aravena, Diego Pasmiño+2 more|arXiv (Cornell University)|Sep 30, 2022

Face recognition and analysis4 citations

TL;DR

This paper investigates how face image quality assessment (FIQA) impacts presentation attack detection (PAD) in face recognition systems. By filtering low-quality bona fide and attack samples before training, a Vision Transformer-based PAD model (ViTranZFAS) achieved a 4% improvement in BPCER on cross-dataset tests when training on a 20%-reduced dataset, demonstrating that quality-aware filtering enhances PAD robustness and generalization.

ABSTRACT

Non-referential face image quality assessment methods have gained popularity as a pre-filtering step on face recognition systems. In most of them, the quality score is usually designed with face matching in mind. However, a small amount of work has been done on measuring their impact and usefulness on Presentation Attack Detection (PAD). In this paper, we study the effect of quality assessment methods on filtering bona fide and attack samples, their impact on PAD systems, and how the performance of such systems is improved when training on a filtered (by quality) dataset. On a Vision Transformer PAD algorithm, a reduction of 20% of the training dataset by removing lower quality samples allowed us to improve the BPCER by 3% in a cross-dataset test.

Motivation & Objective

• To evaluate the impact of face image quality assessment (FIQA) on presentation attack detection (PAD) performance in real-world biometric systems.
• To analyze how FIQA pre-filtering affects the selection of bona fide versus attack samples, particularly in terms of quality distribution.
• To investigate whether removing low-quality samples from the training dataset improves PAD generalization and reduces error rates.
• To compare the effectiveness of different FIQA methods (MagFace, SER-FIQ, SDD-FIQ) in enhancing PAD system performance across multiple datasets.

Proposed method

• Applied multiple state-of-the-art FIQA methods—MagFace, SER-FIQ, and SDD-FIQ—to score and filter training images based on quality.
• Filtered training datasets by discarding 20% of the lowest-quality samples (both bona fide and attack images) to assess impact on PAD model performance.
• Trained and evaluated two PAD models—ViTranZFAS and SSDG—on both full and filtered datasets using intra- and cross-dataset testing protocols.
• Measured performance using standard metrics: EER (Equal Error Rate), BPCER (Between-10% and 20% FAR), and APCER (Attack Presentation Classification Error Rate).
• Used a Vision Transformer-based architecture (ViTranZFAS) as the primary PAD model to evaluate generalization under quality-filtered data.
• Conducted ablation studies to compare performance across different FIQA methods and their influence on PAD accuracy and robustness.

Experimental results

Research questions

• RQ1 How does pre-filtering with FIQA methods affect the balance between bona fide and attack samples in training data?
• RQ2 To what extent does removing low-quality samples improve the generalization of PAD models across different datasets?
• RQ3 Which FIQA method yields the most consistent improvement in PAD performance when used for data filtering?
• RQ4 Does the performance gain from quality filtering vary between intra-dataset and cross-dataset evaluation scenarios?
• RQ5 Can quality-aware filtering reduce the number of false positives in PAD systems without sacrificing bona fide recognition accuracy?

Key findings

• Filtering out 20% of the lowest-quality samples reduced the BPCER by 4% on cross-dataset tests when training the ViTranZFAS model, indicating improved generalization.
• The MagFace FIQA method produced the best overall improvements, reducing EER by up to 1% and BPCER by up to 4% across multiple test scenarios.
• The ViTranZFAS model showed better discrimination performance on high-quality attack samples, suggesting that quality-aware filtering enhances attack detection even for spoofed faces.
• Training with filtered data (20% reduction) improved BPCER by 3% on cross-dataset tests, demonstrating that data quality filtering enhances robustness.
• The SSDG model also benefited from MagFace-based filtering, showing a 1.5% improvement in EER on intra-dataset tests and reduced BPCER on cross-dataset evaluations.
• In contrast, MobileNetv3-based training showed no performance gains and even worsened metrics when using filtered data, indicating model-specific sensitivity to data quality filtering.