[Paper Review] Deep Learning based Fingerprint Presentation Attack Detection: A Comprehensive Survey

Hailin Li, Raghavendra Ramachandra|arXiv (Cornell University)|May 27, 2023

Biometric Identification and Security4 citations

TL;DR

This comprehensive survey reviews deep learning-based fingerprint presentation attack detection (FPAD) methods, categorizing them into contact, contactless, and smartphone-based approaches. It highlights state-of-the-art techniques, identifies key challenges such as interpretability, lightweight modeling, and dataset scarcity, and discusses emerging threats like adversarial attacks, providing a roadmap for future research in secure biometric authentication.

ABSTRACT

The vulnerabilities of fingerprint authentication systems have raised security concerns when adapting them to highly secure access-control applications. Therefore, Fingerprint Presentation Attack Detection (FPAD) methods are essential for ensuring reliable fingerprint authentication. Owing to the lack of generation capacity of traditional handcrafted based approaches, deep learning-based FPAD has become mainstream and has achieved remarkable performance in the past decade. Existing reviews have focused more on hand-cratfed rather than deep learning-based methods, which are outdated. To stimulate future research, we will concentrate only on recent deep-learning-based FPAD methods. In this paper, we first briefly introduce the most common Presentation Attack Instruments (PAIs) and publicly available fingerprint Presentation Attack (PA) datasets. We then describe the existing deep-learning FPAD by categorizing them into contact, contactless, and smartphone-based approaches. Finally, we conclude the paper by discussing the open challenges at the current stage and emphasizing the potential future perspective.

Motivation & Objective

• To provide a current, focused review of deep learning-based fingerprint presentation attack detection (FPAD) methods, moving beyond outdated handcrafted approaches.
• To categorize and analyze recent FPAD techniques across contact, contactless, and smartphone-based modalities.
• To identify open challenges in interpretability, model efficiency, dataset scalability, and emerging threats like adversarial attacks.
• To guide future research by outlining key opportunities and directions in secure biometric authentication.

Proposed method

• Categorizes deep learning-based FPAD methods into three main modalities: contact-based (optical, capacitive, ultrasonic), contactless (industrial and smartphone cameras), and smartphone-based fingerprint capture.
• Reviews state-of-the-art deep learning architectures used in FPAD, emphasizing feature learning from raw or preprocessed fingerprint images.
• Analyzes visualization techniques such as Grad-CAM and saliency maps to improve model interpretability and explain predictions.
• Proposes lightweight model design for mobile deployment, focusing on efficiency and regional feature attention.
• Discusses the use of multi-spectral imaging and advanced sensors to enhance spoof detection robustness.
• Examines adversarial attack vectors, including digital perturbations and physical spoof fabrication (e.g., ScreenSpoof), to assess system vulnerabilities.

Figure 1. Publications of FPAD in recent years obtained through Google scholar search with keywords: “fingerprint spoof detection”, “fingerprint presentation attack detection”, and “fingerprint liveness detection”.

Experimental results

Research questions

• RQ1 How do deep learning-based FPAD methods differ across contact, contactless, and smartphone-based fingerprint acquisition modalities?
• RQ2 What are the key technical challenges in achieving robust, generalizable, and interpretable FPAD models across diverse spoofing materials and sensor types?
• RQ3 How can lightweight deep learning models be designed to support real-time, on-device FPAD in mobile biometric systems?
• RQ4 What are the implications of adversarial attacks on FPAD systems, and how can they be mitigated?
• RQ5 What are the major gaps in publicly available datasets for training and evaluating deep learning-based FPAD systems, especially for contactless finger photos?

Key findings

• Deep learning-based FPAD has become the mainstream approach due to superior performance over traditional handcrafted methods, particularly in handling complex spoofing materials.
• Contactless and smartphone-based FPAD methods face challenges such as pose variation, low image quality, and environmental noise, requiring specialized deep learning architectures.
• Interpretability techniques like Grad-CAM help visualize decision-relevant regions in fingerprint images, improving model transparency and trustworthiness.
• Lightweight models are essential for mobile deployment, where computational resources are limited, and focusing on key fingerprint regions can improve efficiency.
• Large-scale, diverse, publicly available datasets—especially for contactless finger photos—are still lacking, hindering model generalization and benchmarking.
• Adversarial attacks, including physical spoofing using digital perturbations (e.g., ScreenSpoof), pose a significant threat, with high success rates even in one-shot attacks, necessitating robust countermeasures.

Figure 2. Block diagram of fingerprint verification system with PAD