An exploratory study of interpretability for face presentation attack detection

Abstract

Biometric recognition and presentation attack detection (PAD) methods strongly rely on deep learning algorithms. Though often more accurate, these models operate as complex black boxes. Interpretability tools are now being used to delve deeper into the operation of these methods, which is why this work advocates their integration in the PAD scenario. Building upon previous work, a face PAD model based on convolutional neural networks was implemented and evaluated both through traditional PAD metrics and with interpretability tools. An evaluation on the stability of the explanations obtained from testing models with attacks known and unknown in the learning step is made. To overcome the limitations of direct comparison, a suitable representation of the explanations is constructed to quantify how much two explanations differ from each other. From the point of view of interpretability, the results obtained in intra and inter class comparisons led to the conclusion that the presence of more attacks during training has a positive effect in the generalisation and robustness of the models. This is an exploratory study that confirms the urge to establish new approaches in biometrics that incorporate interpretability tools. Moreover, there is a need for methodologies to assess and compare the quality of explanations.This is an open access article under the terms of the Creative Commons Attribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes.

An Attention-Guided Framework for Explainable Biometric Presentation Attack Detection

“…Additionally, the saliency maps are used to further train the classifier of the PAD system, thus enhancing its performance, as shown in Table 3. Compared to the recently published papers [25,26], the work presented here has the additional advantage of producing human-readable explanations.…”

“…Much of the current literature in this area pays particular attention to defining "what is the explanation". Visualization of the filters in a CNN, also referred to as perceptive interpretability methods [14,25,26], is one of the direct ways to explore patterns hidden within the neural units. The Up-convolutional network [27] was developed to reverse the feature map back to an image.…”

“…On the other hand, gradient-based visualization [28] provides a different way of understanding the knowledge hidden within the parameters of a CNN. However, the visual interpretation approaches may generate some meaningless salience maps [25]. There is no commonly used evaluation methodology to quantitatively measure the effectiveness of the visual explanations [26].…”

Deep Learning for Face Anti-Spoofing: A Survey

“…Recently, great efforts have been achieved on interpretable FAS [183]. Some methods try to localize the spoof regions according to the feature activation using visual interpretability tools (e.g., Grad-CAM [184]) or soft-gating strategy [131].…”