Explainable Deep Liveness Detection: Balancing Transparency, Security, and Compliance in Workforce and Government Biometric Systems

Rahul Raj

Abstract: Background: Face anti-spoofing (FAS) systems deployed in workforce management and government identity verification face dual imperatives: achieving robust detection performance across diverse presentation attacks while maintaining transparency to satisfy regulatory requirements such as GDPR Article 22 and ISO/IEC 30107-3. Existing deep learning approaches often operate as black boxes, limiting trust and auditability in high-stakes applications. Recent hybrid architectures combining Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) demonstrate superior cross-domain generalization, yet their decision-making processes remain opaque without appropriate explainability mechanisms. Methods: This study proposes an Explainable Deep Liveness Detection (EDLD) framework integrating a hybrid CNN-Transformer architecture with dual explainability pathways: Layer-wise Relevance Propagation (LRP) for pixel-level artifact localization and SHapley Additive exPlanations (SHAP) for global feature attribution. The framework was evaluated on three benchmark datasets (OULU-NPU, SiW, and a proprietary Workforce-Augmented dataset) comprising 47,832 genuine and spoofed samples across print, replay, and 3D mask attacks. Performance metrics included Average Classification Error Rate (ACER), Attack Presentation Classification Error Rate (APCER), and Bona-fide Presentation Classification Error Rate (BPCER). Explainability was assessed through faithfulness scores (perturbation-based fidelity) and comprehensibility metrics (human operator studies with n=24 security analysts). Results: The EDLD framework achieved state-of-the-art performance with ACER of 2.1% on OULU-NPU Protocol 1, 3.8% on SiW Protocol 2, and 4.2% on the Workforce-Augmented cross-domain protocol, representing 7.3% and 12.9% improvements over standalone CNN and ViT baselines respectively. LRP demonstrated superior pixel-level precision (faithfulness score: 0.87) in localizing morphing artifacts and print attack boundaries, while SHAP excelled in revealing global decision patterns and dataset biases (comprehensibility score: 4.2/5.0 from operator evaluations). The dual explainability approach identified critical model reliance on periocular regions (42% attribution weight) and texture inconsistencies (31% attribution weight), enabling targeted model refinement and bias mitigation. Conclusions: The EDLD framework successfully bridges the performance-transparency gap in face anti-spoofing for workforce and government applications. By combining hybrid architectural advantages with complementary explainability methods, the system achieves both robust cross-domain detection and regulatory-compliant transparency. LRP's pixel-level precision supports forensic analysis and operator training, while SHAP's model-agnostic attribution enables systematic bias auditing and compliance documentation. Future research should address temporal explainability for video-based liveness detection, develop standardized metrics for explanation quality in biometric contexts, and establish compliance-by-design frameworks that integrate ISO/IEC 30107-3 reporting requirements directly into model training pipelines.

Keywords: Face Anti-Spoofing, Explainable AI, Hybrid CNN-Transformer, Layer-wise Relevance Propagation, SHAP, Biometric Security, Regulatory Compliance, Workforce Authentication

How to Cite?: Rahul Raj, "Explainable Deep Liveness Detection: Balancing Transparency, Security, and Compliance in Workforce and Government Biometric Systems", Volume 15 Issue 2, February 2026, International Journal of Science and Research (IJSR), Pages: 1359-1376, https://www.ijsr.net/getabstract.php?paperid=SR26207024823, DOI: https://dx.dx.doi.org/10.21275/SR26207024823

Download Citation: APA | MLA | BibTeX | EndNote | RefMan