Optimizing CNN Architectures for Face Liveness Detection: Performance, Efficiency, and Generalization across Datasets

Face liveness detection is essential for securing biometric authentication systems against spoofing attacks, including printed photos, replay videos, and 3D masks. This study systematically evaluates pre-trained CNN models— DenseNet201, VGG16, InceptionV3, ResNet50, VGG19, MobileNetV2, Xception, and InceptionResNetV2—leveraging transfer learning and fine-tuning to enhance liveness detection performance. The models were trained and tested on NUAA and Replay-Attack datasets, with cross-dataset generalization validated on SiW-MV2 to assess real-world adaptability. Performance was evaluated using accuracy, precision, recall, FAR, FRR, HTER, and specialized spoof detection metrics (APCER, NPCER, ACER). Fine-tuning significantly improved detection accuracy, with DenseNet201 achieving the highest performance (98.5% on NUAA, 97.71% on Replay-Attack), while MobileNetV2 proved the most efficient model for real-time applications (latency: 15 ms, memory usage: 45 MB, energy consumption: 30 mJ). A statistical significance analysis (paired t-tests, confidence intervals) validated these improvements. Cross-dataset experiments identified DenseNet201 and MobileNetV2 as the most generalizable architectures, with DenseNet201 achieving 86.4% accuracy on Replay-Attack when trained on NUAA, demonstrating robust feature extraction and adaptability. In contrast, ResNet50 showed lower generalization capabilities, struggling with dataset variability and complex spoofing attacks. These findings suggest that MobileNetV2 is well-suited for low-power applications, while DenseNet201 is ideal for high-security environments requiring superior accuracy. This research provides a framework for improving real-time face liveness detection, enhancing biometric security, and guiding future advancements in AI-driven anti-spoofing techniques.