Abstract
Face Recognition Systems (FRSs) are increasingly deployed in high-security environments like border control. However, they remain vulnerable to face morphing attacks, where a synthetic image is created by blending the features of two distinct individuals. This paper introduces a novel Quadruplet Loss function designed to improve the robustness of FRSs. By explicitly incorporating morphed images during training and enforcing specific distance constraints, the system learns to differentiate between genuine identities and synthetic blends. Experimental results demonstrate that this approach significantly reduces the success rate of morphing attacks, particularly those generated by modern GAN-based techniques.
Introduction
The widespread adoption of Face Recognition Systems (FRSs) for identity verification has made them a prime target for sophisticated fraud. One of the most challenging threats is the "face morphing" attack. In this scenario, an attacker creates a synthetic face image that contains biometric characteristics of two different people. If this morphed image is used in a passport or ID, both individuals might be able to use that document to pass through automated security gates.
Standard training objectives, such as the Triplet Loss, focus on clustering images of the same identity while separating different identities. However, these methods do not account for the "middle-ground" nature of morphed images. Morphed samples often fall between the feature clusters of the two source identities in the feature space, leading to false matches and system confusion.
Methodology: The Quadruplet Loss
To address the limitations of standard training, the authors propose an evolution of the Triplet Loss into a Quadruplet Loss specifically tailored for morphing detection.
From Triplet to Quadruplet
The standard Triplet Loss utilizes three samples: an Anchor (A), a Positive (P, same identity as A), and a Negative (N, different identity). The objective is to minimize the distance between the Anchor and Positive while maximizing the distance between the Anchor and Negative.
The proposed Quadruplet Loss introduces a fourth element: the Morphed (M) image. The loss function is redesigned to enforce four primary constraints:
- Maintain proximity between the Anchor and Positive samples.
- Ensure a significant margin between the Anchor and Negative samples.
- Crucially, push the Morphed image away from both the Anchor and the Negative identities.
- Ensure the Morphed image is also pushed away from the Positive sample.
By treating the morphed image as a separate entity to be rejected, the network "carves out" tighter boundaries around genuine identity clusters. The authors also investigated "self-morphs"—morphs created from two images of the same person—to determine if the network was learning to detect general synthetic artifacts or specific identity-blending traits.
Experimental Setup and Datasets
The researchers utilized a filtered version of the VGGFace2 dataset for training, ensuring high-quality source images. Evaluation was conducted against two primary types of morphing:
- Landmark-based Morphing (LDM): A traditional approach using geometric landmarks to blend faces.
- StyleGAN-based Morphing (STG): A modern approach using Generative Adversarial Networks to create highly realistic blends that often lack the visual artifacts found in LDM.
The benchmark included both ICAO-compliant (professional camera) and non-compliant (smartphone) images to simulate real-world enrollment and verification conditions.
Results and Discussion
The effectiveness of the approach was measured using the Mismatched Morph Presentation Match Rate (MMPMR), where a lower value indicates better security.
The Quadruplet Loss showed its greatest strength against StyleGAN (STG) attacks. At a fixed False Non-Match Rate (FNMR) of 0.01, the Quadruplet model achieved a MinMax-MMPMR of 0.918, a significant improvement over the Triplet baseline's 0.999. In the case of Landmark (LDM) morphs, the Quadruplet approach maintained high performance without sacrificing the system's baseline accuracy for genuine identity verification.
A key insight from the "self-morph" experiments was that the performance gains were primarily driven by the network's improved ability to separate identities in the feature space, rather than simply detecting the visual glitches (artifacts) common in synthetic images.
Conclusion
The introduction of Quadruplet Loss represents a significant step forward in securing biometric systems against identity-blending attacks. By explicitly modeling the "no-man's-land" occupied by morphed images, the proposed method creates more resilient face recognition models. Future research will focus on the impact of "print and scan" cycles on morph detection and expanding the methodology to handle newer generative models, such as diffusion-based morphing.
Summary updated