<![CDATA[Presentation attacks, commonly known as face spoofing, remain a major security challenge in facial recognition-based authentication systems because forged media such as printed photos and replayed videos can deceive biometric verification. Lightweight CNN models such as MobileNetV2 are suitable for practical implementation, but their limited representational capacity may affect their ability to capture subtle spoofing cues and generalize to unseen identities. Previous evaluations may also produce inflated performance estimates when images from the same identity appear across training and testing sets. This study evaluates Knowledge Distillation and Optuna-based hyperparameter tuning on MobileNetV2 for lightweight face anti-spoofing under an identity-aware evaluation protocol. The novelty lies in an identity-aware comparison between representation enhancement through EfficientNet-B0-based Knowledge Distillation and optimization-based improvement through Optuna. A total of 60,000 CelebA-Spoof images were divided using an 80:10:10 subject-disjoint split, and four scenarios were compared. The baseline MobileNetV2 achieved the best overall balance, with an accuracy of 0.9943, F1-score of 0.9958, and ACER of 0.0058. Meanwhile, Knowledge Distillation obtained the lowest APCER of 0.0035, indicating fewer spoof samples were incorrectly accepted as live under the identity-aware evaluation setting.]]>
Copyrights © 2026
