Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quadruplet Loss For Improving the Robustness to Face Morphing Attacks

Iurii Medvedev, Nuno Gonçalves

TL;DR

The paper tackles the vulnerability of face recognition systems to morphing attacks by introducing a morphing-aware quadruplet loss that extends the traditional triplet framework. It integrates morphed samples into the training objective with weighted cross-sample distances to enhance the discrimination between genuine and morphed representations. Leveraging MorDeephy-derived data, landmark-based morphs, and StyleGAN morphs, the authors demonstrate through benchmarking that their approach yields superior robustness against morphing compared with a Triplet baseline, with self-morphs exerting limited influence. This work offers a practical pathway to more secure biometric verification by strengthening face embeddings against morphing, and it lays groundwork for broader evaluations across morph types and network architectures.

Abstract

Recent advancements in deep learning have revolutionized technology and security measures, necessitating robust identification methods. Biometric approaches, leveraging personalized characteristics, offer a promising solution. However, Face Recognition Systems are vulnerable to sophisticated attacks, notably face morphing techniques, enabling the creation of fraudulent documents. In this study, we introduce a novel quadruplet loss function for increasing the robustness of face recognition systems against morphing attacks. Our approach involves specific sampling of face image quadruplets, combined with face morphs, for network training. Experimental results demonstrate the efficiency of our strategy in improving the robustness of face recognition networks against morphing attacks.

Quadruplet Loss For Improving the Robustness to Face Morphing Attacks

TL;DR

The paper tackles the vulnerability of face recognition systems to morphing attacks by introducing a morphing-aware quadruplet loss that extends the traditional triplet framework. It integrates morphed samples into the training objective with weighted cross-sample distances to enhance the discrimination between genuine and morphed representations. Leveraging MorDeephy-derived data, landmark-based morphs, and StyleGAN morphs, the authors demonstrate through benchmarking that their approach yields superior robustness against morphing compared with a Triplet baseline, with self-morphs exerting limited influence. This work offers a practical pathway to more secure biometric verification by strengthening face embeddings against morphing, and it lays groundwork for broader evaluations across morph types and network architectures.

Abstract

Recent advancements in deep learning have revolutionized technology and security measures, necessitating robust identification methods. Biometric approaches, leveraging personalized characteristics, offer a promising solution. However, Face Recognition Systems are vulnerable to sophisticated attacks, notably face morphing techniques, enabling the creation of fraudulent documents. In this study, we introduce a novel quadruplet loss function for increasing the robustness of face recognition systems against morphing attacks. Our approach involves specific sampling of face image quadruplets, combined with face morphs, for network training. Experimental results demonstrate the efficiency of our strategy in improving the robustness of face recognition networks against morphing attacks.
Paper Structure (13 sections, 5 equations, 4 figures, 1 table)

This paper contains 13 sections, 5 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Face Morphing
  4. Face Recognition
  5. Quadruplet loss
  6. Methodology
  7. Quadruplet Loss With Face Morphing
  8. Data Harvesting
  9. Data Sampling
  10. Experiments
  11. Benchmarking
  12. Results
  13. Conclusion

Figures (4)

  • Figure 1: The proposed Quadruplet Loss minimizes the distance between an anchor and a positive, both of which have the same identity, and maximizes the distance between the anchor and a negative, anchor and morphed, negative and morphed where anchor and negative belongs to different identities and morphed is taken between the positive and negative.
  • Figure 2: The distribution of features $x$ (with two dimensional feature layer) with common face recognition approaches (left) and desired feature distribution in proposed approach (right). Each point on 2D surface corresponds to a single image features (data is not based on real experiments).
  • Figure 3: Quadruplet samples
  • Figure 4: MinMax-MMPMR and ProdAvg-MMPMR for different models and protocols as a dependency from FNMR. a,b - LDM benchmark; c,d - STG benchmark.