Deep Guard: An enhanced hybrid ensemble classifier for face presentation attack detection integrating Gabor and binarized statistical image features descriptors with deep learning

Facial recognition systems are widely used in various real-world applications due to their reliability and convenience. However, attackers exploit these systems by mimicking bona fide user traits to gain unauthorized access. This emphasizes the need for effective countermeasures to be integrated into face-based authentication systems. Face presentation attack detection methods encounter several challenges such as illumination variations and noisy input images which limit the performance of the attack detection methods, particularly on unseen data. In this paper, we introduce Deep Guard, a hybrid framework that combines handcrafted texture descriptors with advanced deep learning techniques. The framework utilizes an ensemble of different classifiers to leverage their complementary strengths. The first classifier applies Binarized Statistical Image Features (BSIF) and a Multilayer Perceptron (MLP) to capture fine-grained texture details. The second classifier combines EfficientNet-B0 with ConvMixer layers and a CBAM attention mechanism to enhance feature representation and improve perceptual capabilities. The third classifier uses Gabor filters as convolutional layers with a deep network which is used in second classifier to refine edges and increase robustness to illumination and noise. The outputs from these classifiers are fused using a soft voting mechanism to classify facial images as real or fake. We evaluate the proposed framework on six publicly available datasets CASIA-FASD, Replay-Attack, 3DMAD, ROSE-Youtu, OULU-NPU, and MSU-MFSD. The results demonstrate that Deep Guard outperforms most state-of-the-art methods in intra-dataset testing and achieves strong generalization performance in cross-dataset single source training and testing scenarios, with an average HTER of 25.78% for HybridNet I, which combines all three classifiers and 27.96% for HybridNet II, combining classifiers two and three. It also achieves an AUC of 98.65% for cross-dataset evaluation with multiple-source training and single-source testing (O&C&I $\to$ M).