Beyond Intra-Dataset Accuracy: Cross-Dataset Generalization of Scratch-Trained and Pretrained CNNs for Face Presentation Attack Detection
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171089 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Nawal Alsufyani.
Abstract (Originalsprache)
Face presentation attack detection (PAD) systems often achieve strong performance under intra-dataset evaluation but experience substantial degradation when exposed to unseen acquisition conditions. This study presents an empirical investigation of the discrepancy between intra-dataset performance and cross-dataset generalization by comparing a lightweight convolutional neural network trained from scratch (SimpleCNN) with an ImageNet-pretrained ResNet18. Both models were trained and evaluated using OULU-NPU, followed by strict zero-shot evaluation on CASIA-FASD without target-domain fine-tuning, adaptation, or threshold recalibration. On OULU-NPU, ResNet18 achieved an Average Classification Error Rate (ACER) of 10.97%, an Equal Error Rate (EER) of 10.22%, and an Area Under the ROC Curve (AUC) of 0.967, outperforming SimpleCNN, which achieved an ACER of 42.67%, an EER of 17.85%, and an AUC of 0.894. However, both models experienced substantial degradation on CASIA-FASD, with ACER values of 50.80% and 50.43% for SimpleCNN and ResNet18, respectively. The corresponding AUC values decreased to 0.403 and 0.305, accompanied by extremely high bona fide presentation classification error rates. These findings reveal substantial limitations in the transferability of the evaluated source-trained classifiers under the investigated cross-dataset protocol. Since the compared configurations differ in both architecture and initialization, the results do not isolate the independent contribution of ImageNet pretraining. Furthermore, the evaluation is limited to one transfer direction and frame-level predictions. The study highlights the importance of rigorous cross-dataset assessment, careful interpretation of class-specific errors, and broader experimental validation when evaluating the robustness of face PAD systems.
Zitieren
APA
Alsufyani, N. (2026). Beyond Intra-Dataset Accuracy: Cross-Dataset Generalization of Scratch-Trained and Pretrained CNNs for Face Presentation Attack Detection. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171089
Vancouver
Alsufyani N. Beyond Intra-Dataset Accuracy: Cross-Dataset Generalization of Scratch-Trained and Pretrained CNNs for Face Presentation Attack Detection. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171089
BibTeX
@article{Alsufyani2026,
author = {Alsufyani, Nawal},
title = {Beyond Intra-Dataset Accuracy: Cross-Dataset Generalization of Scratch-Trained and Pretrained CNNs for Face Presentation Attack Detection},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171089}
}
EUVIMED