Domain Generalization Performance of CNN Architectures for Face Anti-Spoofing Using a Multi-Dataset Evaluation

Abstract

Face recognition is widely used to improve the security level and digital attendance systems, but even still, it can susceptible to the spoofing using photos, videos or fake mask. The objective of this work is to evaluate the generalization performance of three CNN architectures (MobileNetV2, EfficientNetB0 and ResNet50) on face spoof detection by applying them into intra-dataset and cross-dataset experiments with SiW-Mv2, Replay-Attack and Paper-Attack datasets. The working of the deep-learning model is explained along with pre-processing of all datasets: firstly, splitting videos into training, validation and test sets. Secondly, frame extraction Thirdly, data augmentation for the training images; and fourthly Normalization-based preprocess_input method. The models use a fine-tuning method to train on the last 30\% of layers; maximum number of epochs is set at 30, batch size typically is {32},and hyperparameters are fixed. Experimental results are evaluated with respect to accuracy, precision, recall, F1-score, confusion matrix and average drop rate as measure of sensitivity of generalization. This experiment confirms that EfficientNetB0 has the lowest performance loss on all cross-dataset scenarios, therefore being the most stable architecture for domain variations. The adjusted model is fine-tuned by optimization of the learning rate and patience value, which performs better in testing on SiW-Mv2. The chosen model is finally tested on facial recognition based attendance systems to demonstrate practical applicability. The reported results lead to the establishment of more secure attendance systems, with higher robustness and less vulnerability to spoof attacks