Are Explainability Tools Gender Biased? A Case Study on Face Presentation Attack Detection

A typical automated face recognition system is composed of three main component tasks: face detection and alignment (FDA), face presentation attack detection (FPAD), and face representation and matching (FRM). These tasks are often treated as standalone problems and deep neural network (DNN)-based solutions have been proposed to address them individually. However, in resource-constrained sce-narios it would be ideal to have a unified DNN model that can perform all the three tasks together. As a first step towards realizing this goal, this work attempts to perform joint FRM and FPAD based on a single Vision Transformer (ViT) backbone. Recent work demonstrating the ability of ViT to extract a diverse set of feature representations gives rise to the tantalising possibility of building an end-to-end face recognition system using a single ViT model. The standard approach for designing multi-task DNNs is to implement different classification heads (e.g., for FRM and FPAD) based on a common stem/base and learn these heads either individually or jointly. A key contribution of this work is to demonstrate that this naive multi-head approach results in sub-optimal performance for either FRM or FPAD, because the features required by these tasks are very different. While good FPAD performance depends on accurately characterizing the micro textures, face matching demands attention towards more global characteris-tics. Hence, we propose a novel feature ensemble approach, where an ensemble of local features extracted from the in-termediate blocks of a ViT are utilized for FPAD, while face matching is performed based on the ViT class token. Exper-iments demonstrate that the proposed ViT feature ensemble approach is able to achieve good performance for both face matching and FPAD compared to the multi-head approach.

State-of-the-art spoof detection methods tend to overfit to the spoof types seen during training and fail to generalize to unknown spoof types. Given that face anti-spoofing is inherently a local task, we propose a face anti-spoofing framework, namely Self-Supervised Regional Fully Convolutional Network (SSR-FCN), that is trained to learn local discriminative cues from a face image in a self-supervised manner. The proposed framework improves generalizability while maintaining the computational efficiency of holistic face anti-spoofing approaches (< 4 ms on a Nvidia GTX 1080Ti GPU). The proposed method is interpretable since it localizes which parts of the face are labeled as spoofs. Experimental results show that SSR-FCN can achieve TDR = 65% @ 2.0% FDR when evaluated on a dataset comprising of 13 different spoof types under unknown attacks while achieving competitive performances under standard benchmark datasets (Oulu-NPU, CASIA-MFSD, and Replay-Attack).

Face presentation attack detection, also termed Face Anti-Spoofing (FAS) [item 1), 2) in the Appendix), is a hot and challenging research topic that has received much attention from the computer vision and pattern recognition communities in the past. Owing to the development of deep learning and big data, recent advances in this and related fields has increased considerably. However, there are still several challenging tasks that deserve attention from the community, for instance robust techniques to unknown spoofing attacks, cross-domain generalization, and multi-modal fusion in images and video sequences. We edited this special issue with the goal of compiling the latest progress in the field and identifying promising research opportunities on FAS.

Face presentation attack detection (PAD) plays a vital role in face recognition systems. Many previous face antispoofing methods mainly focus on the 2D face representation attacks, which however, suffer from great performance degradation when facing high-fidelity 3D mask attacks. To address this issue, we propose a novel dual-stream framework consisting of the vanilla convolution stream and the central difference convolution stream. These two streams complement each other and learn more comprehensive features for 3D mask attacks detection. Moreover, we extend 3D PAD to a multi-classification task that contains real face, plaster attack and transparent attack, and utilize various data augmentations and label smoothing techniques to improve the generalizability on unseen attacks. The proposed method achieved the second place in the Chalearn 3D High-Fidelity Mask Face Presentation Attack Detection Challenge@ICCV2021 with a score of 3.15 (ACER).

Face-based biometric recognition systems that can recognize human faces are widely employed in places such as airports, immigration offices, and companies, and applications such as mobile phones. However, the security of this recognition method can be compromised by attackers (unauthorized persons), who might bypass the recognition system using artificial facial images. In addition, most previous studies on face presentation attack detection have only utilized spatial information. To address this problem, we propose a visible-light camera sensor-based presentation attack detection that is based on both spatial and temporal information, using the deep features extracted by a stacked convolutional neural network (CNN)-recurrent neural network (RNN) along with handcrafted features. Through experiments using two public datasets, we demonstrate that the temporal information is sufficient for detecting attacks using face images. In addition, it is established that the handcrafted image features efficiently enhance the detection performance of deep features, and the proposed method outperforms previous methods.

Detecting anomalies using deep learning has become a major challenge over the last years, and is becoming increasingly promising in several fields. The introduction of self-supervised learning has greatly helped many methods including anomaly detection where simple geometric transformation recognition tasks are used. However these methods do not perform well on fine-grained problems since they lack finer features. By combining both high-scale shape features and low-scale fine features in a multi-task framework, our method greatly improves fine-grained anomaly detection. It outperforms state-of-the-art with up to 31% relative error reduction measured with AUROC on various anomaly detection problems including one-vs-all, out-of-distribution detection and face presentation attack detection.

Deepfake-enabled identity abuse has moved from a peripheral cyber-risk to an operational threat for U.S. banking, especially in remote onboarding, account recovery, contact-center authentication, and high-risk payment authorization. The problem is no longer limited to obvious synthetic media. Financial institutions increasingly face blended attacks that combine forged identity documents, face or voice cloning, injected video streams, social-engineering pressure, mule accounts, and adaptive retries across channels. This paper develops a publication-ready research framework for machine learning systems that identify deepfake-driven customer impersonation, authentication evasion, and identity abuse in U.S. banking environments. The study synthesizes regulatory guidance, public fraud data, biometric spoofing research, deepfake detection literature, and banking fraud analytics to propose a multimodal detection architecture spanning document forensics, face presentation attack detection, deepfake-video detection, speaker anti-spoofing, behavioral biometrics, device and network telemetry, graph-based entity resolution, and risk-calibrated decisioning. Real public evidence motivates the design: FTC data show reported fraud losses reached $12.5 billion in 2024, including $2.95 billion in imposter-scam losses, while FinCEN reported increasing suspicious activity narratives involving deepfake media targeting financial institutions. The proposed methodology treats identity abuse as a sequential, multimodal, and adversarial classification problem rather than a single-screen biometric check. The paper argues that the most effective defense is not one detector but an explainable ensemble with smart friction, human escalation, and governance aligned to NIST identity-proofing standards, anti-money-laundering expectations, and consumer-protection obligations. By linking technical detection with operational controls, the study provides a practical blueprint for authentication, resilient fraud prevention, and more trustworthy digital banking systems.

Today, face biometric systems are becoming widely accepted as a standard method for identity authentication in many security settings. For example, their deployment in automated border control gates plays a crucial role in accurate document authentication and reduced traveler flow rates in congested border zones. The proliferation of such systems is further spurred by the advent of portable devices. On the one hand, modern smartphone and tablet cameras have in-built user authentication applications while on the other hand, their displays are being consistently exploited for face spoofing. Similar to biometric systems of other physiological biometric identifiers, face biometric systems have their own unique set of potential vulnerabilities. In this work, these vulnerabilities (presentation attacks) are being explored via a biologically-inspired presentation attack detection model which is termed “BIOPAD.” Our model employs Gabor features in a feedforward hierarchical structure of layers that progressively process and train from visual information of people's faces, along with their presentation attacks, in the visible and near-infrared spectral regions. BIOPAD's performance is directly compared with other popular biologically-inspired layered models such as the “Hierarchical Model And X” (HMAX) that applies similar handcrafted features, and Convolutional Neural Networks (CNN) that discover low-level features through stochastic descent training. BIOPAD shows superior performance to both HMAX and CNN in all of the three presentation attack databases examined and these results were consistent in two different classifiers (Support Vector Machine and k-nearest neighbor). In certain cases, our findings have shown that BIOPAD can produce authentication rates with 99% accuracy. Finally, we further introduce a new presentation attack database with visible and near-infrared information for direct comparisons. Overall, BIOPAD's operation, which is to fuse information from different spectral bands at both feature and score levels for the purpose of face presentation attack detection, has never been attempted before with a biologically-inspired algorithm. Obtained detection rates are promising and confirm that near-infrared visual information significantly assists in overcoming presentation attacks.

Face recognition systems face real challenges from various presentation attacks. New, more sophisticated methods of presentation attacks are becoming more difficult to detect using traditional face recognition systems. Thermal infrared imaging offers specific physical properties that may boost presentation attack detection capabilities. The aim of this paper is to present outcomes of investigations on the detection of various face presentation attacks in thermal infrared in various conditions including thermal heating of masks and various states of subjects. A thorough analysis of presentation attacks using printed and displayed facial photographs, 3D-printed, custom flexible 3D-latex and silicone masks is provided. The paper presents the intensity analysis of thermal energy distribution for specific facial landmarks during long-lasting experiments. Thermalization impact, as well as varying the subject’s state due to physical effort on presentation attack detection are investigated. A new thermal face spoofing dataset is introduced. Finally, a two-step deep learning-based method for the detection of presentation attacks is presented. Validation results of a set of deep learning methods across various presentation attack instruments are presented.

The main scope of this chapter is to serve as an introduction to face presentation attack detection, including key resources and advances in the field in the last few years. The next pages present the different presentation attacks that a face recognition system can confront, in which an attacker presents to the sensor, mainly a camera, a Presentation Attack Instrument (PAI), that is generally a photograph, a video, or a mask, with the target to impersonate a genuine user or to hide the actual identity of the attacker via obfuscation. First, we make an introduction of the current status of face recognition, its level of deployment, and its challenges. In addition, we present the vulnerabilities and the possible attacks that a face recognition system may be exposed to, showing that way the high importance of presentation attack detection methods. We review different types of presentation attack methods, from simpler to more complex ones, and in which cases they could be effective. Then, we summarize the most popular presentation attack detection methods to deal with these attacks. Finally, we introduce public datasets used by the research community for exploring vulnerabilities of face biometrics to presentation attacks and developing effective countermeasures against known PAIs.

The vulnerability of current face recognition systems to presentation attacks significantly limits their application in biometrics. Herein, we present a passive presentation attack detection method based on a complete plenoptic imaging system which can derive the complete plenoptic function of light rays using a single detector. Moreover, we constructed a multi-dimensional face database with 50 subjects and seven different types of presentation attacks. We experimentally demonstrated that our approach outperforms the state-of-the-art methods on all types of presentation attacks.

Presentation attacks are becoming a serious threat to one of the most common biometric applications, namely face recognition (FR). In recent years, numerous methods have been presented to detect and identify these attacks using publicly available datasets. However, such datasets are often collected in controlled environments and are focused on one specific type of attack. We hypothesise that a model’s accurate performance on one or more public datasets does not necessarily guarantee generalisation across other, unseen face presentation attacks. To verify our hypothesis, in this paper, we present an experimental framework where the generalisation ability of pre-trained deep models is assessed using four popular and commonly used public datasets. Extensive experiments were carried out using various combinations of these datasets. Results show that, in some circumstances, a slight improvement in model performance can be achieved by combining different datasets for training purposes. However, even with a combination of public datasets, models still could not be trained to generalise to unseen attacks. Moreover, models could not necessarily generalise to a learned format of attack over different datasets. The work and results presented in this paper suggest that more diverse datasets are needed to drive this research as well as the need for devising new methods capable of extracting spoof-specific features which are independent of specific datasets.

Compared to 2D face presentation attacks (e.g. printed photos and video replays), 3D type attacks are more challenging to face recognition systems (FRS) by presenting 3D characteristics or materials similar to real faces. Existing 3D face spoofing databases, however, mostly based on 3D masks, are restricted to small data size or poor authenticity due to the production difficulty and high cost. In this work, we introduce the first wax figure face database, WFFD, as one type of super-realistic 3D presentation attacks to spoof the FRS. This database consists of 2200 images with both real and wax figure faces (totally 4400 faces) with a high diversity from online collections. Experiments on this database first investigate the vulnerability of three popular FRS to this kind of new attack. Further, we evaluate the performance of several face presentation attack detection methods to show the attack abilities of this super-realistic face spoofing database.

Light Field Focal Stack (LFFS) can be efficiently rendered from a light field (LF) image captured by plenoptic cameras. Differences in the 3D surface and texture of biometric samples are internally reflected in the defocus blur and local patterns between the rendered slices of LFFS. This unique property makes LFFS quite appropriate to differentiate presentation attack instruments (PAIs) from bona fide samples. A patch-based dual-view network (PDVN) is proposed in this paper to leverage the merits of LFFS for face presentation attack detection (PAD). First, original LFFS data are divided into various local patches along spatial dimensions, which distracts the model from learning the useless facial semantics and greatly relieve the problem of insufficient samples. The strategy of dual-view branches is innovatively proposed, wherein the original view and microscopic view can simultaneously contribute to liveness detection. Separable 3D convolution on the focal dimension is verified to be more effective than vanilla 3D convolution for extracting discriminative features from LFFS data. The voting mechanism on predictions of patch LFFS samples further strengthens the robustness of the proposed framework. PDVN is compared with other face PAD methods on IST LLFFSD dataset and achieves perfect performance, i.e., ACER drops to 0.

Face presentation attack detection (PAD) has received increasing attention ever since the vulnerabilities to spoofing have been widely recognized. The state of the art in unimodal and multi-modal face anti-spoofing has been assessed in eight international competitions organized in conjunction with major biometrics and computer vision conferences in 2011, 2013, 2017, 2019, 2020 and 2021, each introducing new challenges to the research community. In this chapter, we present the design and results of the five latest competitions from 2019 until 2021. The first two challenges aimed at evaluating the effectiveness of face PAD in multi-modal setup introducing near-infrared (NIR) and depth modalities in addition to colour camera data, while the latest three competitions focused on evaluating domain and attack type generalization abilities of face PAD algorithms operating on conventional colour images and videos. We also discuss the lessons learnt from the competitions and future challenges in the field in general.