Explainable Face Presentation Attack Detection via Ensemble-CAM

TL;DR

This work addresses the lack of explainability in deep learning–based face presentation attack detection (PAD) by introducing Ensemble-CAM, a visual explanation method that fuses Grad-CAM, HiResCAM, and Grad-CAM++ to produce highly focused, discriminative heatmaps of model decisions. The approach computes $Avg\_CAM = \frac{Grad\_CAM + HiResCAM + Grad\_CAM++}{3}$ and derives $Ensemble\_CAM$ via thresholding at the $90^{th}$ percentile, yielding more robust localization than any single CAM technique. The DenseNet-161 PAD model is trained on CelebA-Spoof and evaluated with APCER $=12.4\%$, BPCER $=0.95\%$, and accuracy $93.33\%$, with Ensemble-CAM providing clearer explanations and improved stability under retention and deletion evaluations. The method enhances transparency and trust in face PAD systems and offers a practical path for diagnosing vulnerabilities and guiding future improvements, with potential extension to other biometric modalities such as fingerprints and irises. $Avg\_CAM = \frac{Grad\_CAM + HiResCAM + Grad\_CAM++}{3}$, followed by $Ensemble\_CAM = Apply\_Threshold(Avg\_CAM)$, captures complementary cues across CAM variants to produce more precise visual explanations.

Abstract

Presentation attacks represent a critical security threat where adversaries use fake biometric data, such as face, fingerprint, or iris images, to gain unauthorized access to protected systems. Various presentation attack detection (PAD) systems have been designed leveraging deep learning (DL) models to mitigate this type of threat. Despite their effectiveness, most of the DL models function as black boxes - their decisions are opaque to their users. The purpose of explainability techniques is to provide detailed information about the reason behind the behavior or decision of DL models. In particular, visual explanation is necessary to better understand the decisions or predictions of DL-based PAD systems and determine the key regions due to which a biometric image is considered real or fake by the system. In this work, a novel technique, Ensemble-CAM, is proposed for providing visual explanations for the decisions made by deep learning-based face PAD systems. Our goal is to improve DL-based face PAD systems by providing a better understanding of their behavior. Our provided visual explanations will enhance the transparency and trustworthiness of DL-based face PAD systems.

Figures (7)

• Figure 1: Examples of presentation attacks on face recognition systems, such as 3D mask (printed and silicon), face printout (with eyeholes), replay attack (mobile display) purnapatra2021face.
• Figure 2: An overview of the Ensemble-CAM method. For the input image, three CAMs are generated using the Grad-CAM, HiResCAM, and Grad-CAM++ methods. Then, an average CAM is computed to combine all the features of these three CAMs. The Ensemble-CAM is generated by applying a threshold to the average CAM.
• Figure 3: Visual explanations of face PAD model predictions. Here, two spoof test images are considered. The model predicts that these are spoof images (correct prediction). From the left, the first image is the original spoof image. Grad-CAM, HiResCAM, and Grad-CAM++ are generated using the PAD model for the predicted class and overlaid on the test image. The red, yellow, and green regions highlight the relevant features, while the blue regions highlight the non-relevant features. The last image in the row shows the Ensemble-CAM overlaid on the test image, highlighting the most important features (for the correct prediction by the model) very specifically and accurately.
• Figure 4: Visual explanations of face PAD model predictions. Here, two live test images are considered. The model predicts that these are live images (correct prediction). From the left, the first image is the original live image. Grad-CAM, HiResCAM, and Grad-CAM++ are generated using the PAD model for the predicted class and overlaid on the test image. The red, yellow, and green regions highlight the relevant features, while the blue regions highlight the non-relevant features. The last image in the row shows the Ensemble-CAM overlaid on the test image, highlighting the most important features (for the correct prediction by the model) very specifically and accurately.
• Figure 5: Visual explanation of wrong prediction made by the face PAD model. Here, a spoof test image is considered. The model predicts that this is a live image (wrong prediction). The top row shows the CAMs for the original class (spoof), and the bottom row shows the CAMs for the predicted class (live). The Ensemble-CAM of the original class highlights the upper part of the face (nose, eyes, and forehead). The Ensemble-CAM of the wrongly predicted class highlights the right eye, right forehead, and the left side of the body; the model fails and makes a wrong prediction due to these features.