Abstract: Face recognition systems are widely deployed for biometric authentication. Despite this, it is well-known that, without any safeguards, face recognition systems are highly vulnerable to presentation attacks. In response to this security issue, several promising methods for detecting presentation attacks have been proposed which show high performance on existing benchmarks. However, an ongoing challenge is the generalization of presentation attack detection methods to unseen and new attack types. To this end, we propose a new T-shirt Face Presentation Attack (TFPA) database of 1,608 T-shirt attacks using 100 unique presentation attack instruments. In an extensive evaluation, we show that this type of attack can compromise the security of face recognition systems and that some state-of-the-art attack detection mechanisms trained on popular benchmarks fail to robustly generalize to the new attacks. Further, we propose three new methods for detecting T-shirt attack images, one which relies on the statistical differences between depth maps of bona fide images and T-shirt attacks, an anomaly detection approach trained on features only extracted from bona fide RGB images, and a fusion approach which achieves competitive detection performance.
Introduction: The T-shirt Attack Threat
Face recognition (FR) systems have become a cornerstone of modern biometric authentication. However, their security is often undermined by Presentation Attacks (PAs), where an adversary uses a spoofing instrument—such as a printed photo or a digital screen—to deceive the system. While traditional research has focused on rigid masks or high-resolution replays, a novel and easily concealed threat has emerged: the T-shirt attack.
T-shirts with faces printed on them represent a unique security challenge. They are inexpensive, non-rigid, and can be worn inconspicuously under other clothing until the moment of the attack. Because the fabric drapes over the human body, it creates subtle depth variations that can bypass simple 3D sensors designed to detect flat paper prints. This paper introduces the first dedicated T-shirt Face Presentation Attack (TFPA) database and evaluates the vulnerability of current systems to this "hiding in plain sight" strategy.
Methodology and the TFPA Database
To rigorously assess this threat, the researchers developed the TFPA database. Using StyleGAN, they generated 100 high-quality synthetic identities, which were then printed onto T-shirts. Eight subjects wore these shirts across various scenarios, including different head poses, stretching the fabric, and concealing the attacker's real face with a mask or hands.
The database contains 1,608 attack images captured in both the visible (RGB) and depth spectra using an Intel RealSense D435 camera. This multi-modal approach is critical, as the primary defense against such attacks lies in the discrepancy between a 2D print and a 3D human head.
Vulnerability Assessment
The study found that standard face detectors, such as MTCNN and RetinaFace, are ironically "too effective." They locate the face on the T-shirt with high confidence, passing it to the recognition engine as a valid input. Consequently, state-of-the-art recognizers like ArcFace and commercial-off-the-shelf (COTS) systems proved highly vulnerable. When the attacker's real face was concealed, the Impostor Attack Presentation Match Rate (IAPMR) exceeded 92.6%, meaning the systems almost always accepted the T-shirt face as the target identity.
Proposed Detection Strategies
Existing Presentation Attack Detection (PAD) systems often fail to generalize to T-shirt attacks because they were trained on different materials like silicone or paper. To bridge this "generalization gap," the authors proposed three new detection methods:
- Depth Variance (DV): A statistical approach that analyzes depth values at specific facial landmarks. While a T-shirt face looks realistic in 2D, its depth profile is significantly flatter than a real human face.
- Anomaly Detection (AD): A model trained exclusively on "bona fide" (real) RGB images using Vision Transformer (ViT) features and a one-class SVM. This method identifies the T-shirt prints as outliers.
- Fusion: A score-level combination of the DV and AD approaches.
The proposed Fusion method achieved a Detection Equal Error Rate (D-EER) of 12.52%, significantly outperforming existing state-of-the-art models which showed error rates as high as 26% on this new dataset.
Conclusion
The T-shirt attack highlights a critical blind spot in current face recognition security. By exploiting the high performance of face detectors and the non-rigid nature of fabric, attackers can easily spoof sophisticated systems. The researchers demonstrate that a multi-modal defense—combining RGB anomaly detection with depth statistics—is the most robust way to counter these evolving threats. Future work will focus on "context-aware" detection, looking beyond the face to the neck and shoulders to identify the transition between a human body and a printed garment.
Summary updated