Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Impact of Synthetic Images on Morphing Attack Detection Using a Siamese Network

Juan Tapia, Christoph Busch

TL;DR

Results show that MAD trained on EfficientNetB0 from FERET, FRGCv2, and FRLL can reach a lower error rate in comparison with SOTA, and that a mixed approach (synthetic + digital) database may help to improve MAD and reduce the error rate.

Abstract

This paper evaluated the impact of synthetic images on Morphing Attack Detection (MAD) using a Siamese network with a semi-hard-loss function. Intra and cross-dataset evaluations were performed to measure synthetic image generalisation capabilities using a cross-dataset for evaluation. Three different pre-trained networks were used as feature extractors from traditional MobileNetV2, MobileNetV3 and EfficientNetB0. Our results show that MAD trained on EfficientNetB0 from FERET, FRGCv2, and FRLL can reach a lower error rate in comparison with SOTA. Conversely, worse performances were reached when the system was trained only with synthetic images. A mixed approach (synthetic + digital) database may help to improve MAD and reduce the error rate. This fact shows that we still need to keep going with our efforts to include synthetic images in the training process.

Impact of Synthetic Images on Morphing Attack Detection Using a Siamese Network

TL;DR

Results show that MAD trained on EfficientNetB0 from FERET, FRGCv2, and FRLL can reach a lower error rate in comparison with SOTA, and that a mixed approach (synthetic + digital) database may help to improve MAD and reduce the error rate.

Abstract

This paper evaluated the impact of synthetic images on Morphing Attack Detection (MAD) using a Siamese network with a semi-hard-loss function. Intra and cross-dataset evaluations were performed to measure synthetic image generalisation capabilities using a cross-dataset for evaluation. Three different pre-trained networks were used as feature extractors from traditional MobileNetV2, MobileNetV3 and EfficientNetB0. Our results show that MAD trained on EfficientNetB0 from FERET, FRGCv2, and FRLL can reach a lower error rate in comparison with SOTA. Conversely, worse performances were reached when the system was trained only with synthetic images. A mixed approach (synthetic + digital) database may help to improve MAD and reduce the error rate. This fact shows that we still need to keep going with our efforts to include synthetic images in the training process.
Paper Structure (15 sections, 4 equations, 6 figures, 2 tables)

This paper contains 15 sections, 4 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Database
  4. SYN-MAD-2022
  5. SOTA Databases
  6. Metrics
  7. Method
  8. Triplet Loss
  9. Experiments and Results
  10. Exp1 - Train SYN-MAD/TEST SDD-Bechmark
  11. Exp2 - Train SYN-MAD/TEST SOTA
  12. Exp3 - Train SOTA/TEST SDD-Benchmark
  13. Exp4 - Train Mix/TEST SDD-Benchmark
  14. Conclusions
  15. Acknowledgements

Figures (6)

  • Figure 1: Example of images from all the databases. a) SDD: Left to Right: Bona fide (Subject-1) synthetic, Morph Synthetic (Subject-2), MIPGAN-I, MIPGAN-II, OpenCV-Morpher, WebMorpher. B) FERET: Left to right: Bona fide (Subject-1), Bona fide (Subject-2), FaceMorpher, FaceFusion, OpenCV-Morpher, UBO-Morpher. c) FRGCv2: Left to right: Bona fide (Subject-1), Bona fide (Subject-2), FaceMorpher, FaceFusion, OpenCV-Morpher, UBO-Morpher. d) FRLL: Left to Right: Bona fide (Subject-1), Bona fide (Subject-2), AMSL, FaceMorpher, OpenCV-Morpher, and StyleGAN.
  • Figure 2: Siamese network.
  • Figure 3: T-SNE distibution of SDD-test benchmark. Each colour represents a morphing tool. Bona fide (test-0), WebMorpher (test-1), FaceMorpher (test-2), MIPGAN-I (test-3), MIPGAN-II (test-4), OpenCV-Morpher (test-5). Black represents the bona fide validation set (val-0), and Blue (start) represents the morph validation set (val-1).
  • Figure 4: Exp1: DET curves belong to each class of SDD-Test benchmark. Left to Right: FaceMorpher, MIPGAN-I, MIPGAN-II, OpenCV-Morpher, WebMorpher. Dot-line indicates BPCER10 and BPCER20, respectively. The EER is reported in parentheses in percentages.
  • Figure 5: Exp4: DET curves belong to each class of SDD-Test benchmark, trained on mix-database. Left to Right: FaceMorpher, MIPGAN-I, MIPGAN-II, OpenCV-Morpher, WebMorpher. Dot-line indicates BPCER10 and BPCER20, respectively. The EER is reported in percentages in parentheses.
  • ...and 1 more figures