Psychophysically Guided Training for Fingerprint Presentation Attack Detection

Abstract

Psychophysically-guided training, which directs model learning based on human-perceptual annotations, has demonstrated generalization improvements across various vision tasks, including biometric presentation attack detection (PAD). In saliency-guided training, model’s decision-making is aligned to image regions marked by human annotators. In perceptually-weighted training, human psychophysical measurements weigh per-sample loss, prioritizing samples with greater ‘difficulty’. This paper presents the first application of both strategies to fingerprint PAD. We conducted a 50-participant study to perceptually annotate 800 fingerprint samples, producing human-sourced saliency maps and annotation-duration-based perceptual weighting coefficients. These are utilized alongside algorithmically-generated “pseudosaliency” (minutiae-based, quality-based, and autoencoder-generated) and additional perceptual weighting coefficients derived from minutiae count and sample quality. Evaluating on the 2021 Fingerprint Liveness Detection Competition testing set, we explore various configurations within nine distinct training scenarios to assess the impact of psychophysically-guided training on accuracy and generalization. Our findings demonstrate the effectiveness of psychophysically-guided training for fingerprint PAD in both limited and large data contexts, and we present multiple configurations capable of earning the first place on the LivDet-2021 benchmark. Our results highlight psychophysically-guided training’s promise for increased model generalization capabilities, its effectiveness when data is limited, and its potential to scale to larger datasets in fingerprint PAD. All collected saliency maps, perceptual weighting coefficient sets, and trained models are released with the paper to support reproducible research.