Exploration of Incremental Synthetic Non-Morphed Images for Single Morphing Attack Detection

AI summary

Overview

• Research area: Biometric security / face morphing attack detection (S-MAD), with a focus on synthetic training data and cross-dataset generalization.
• Technical level: Intermediate — assumes familiarity with CNNs, transfer learning, and biometric error-rate metrics such as D-EER and BPCER.
• Scope: The paper evaluates, through an incremental testing protocol, how adding synthetic "non-morphed" face images to training affects Single Morphing Attack Detection performance across two cross-dataset evaluation schemes.

What This Paper Is About

Face morphing attacks combine images of two or more people into one image that can be used to obtain fraudulent identity documents, so automated detectors are needed. Training these detectors is difficult because privacy rules limit access to large sets of genuine (bona fide) face images. The authors test whether synthetic non-morphed images from the SMDD dataset can fill that gap, adding them incrementally to real-data training sets and measuring what happens to detection accuracy.

Key Contributions

1. An evaluation of the impact of integrating synthetically generated "non-morphed" images into the training process as a substitute for scarce bona fide images, motivated by privacy constraints on deep learning training data.
2. An incremental testing protocol in which synthetic non-morphed images are added to the training set in fixed proportions (10%, 20%, 30%, 50%, 75% and 100%) to observe the effect on generalization as the synthetic share grows.
3. A cross-dataset evaluation using two morphing tools sets and two round-trip schemes (train FERET / test FRGCv2 and train FRGCv2 / test FERET), with morphed subsets generated by FaceFusion, Face Morpher, OpenCV and UBO.
4. A comparison of two compact backbones, MobileNetV3-large (around 7.7 million trainable parameters) and EfficientNet-B2 (2.9M trainable parameters), intended to reflect realistic deployment constraints such as border control and mobile identity verification.

Main Findings

• Synthetic-only training is the worst option: training exclusively on SMDD morphed and non-morphed data produced D-EERs of 37.96% (EfficientNet-B2) and 38.95% (MobileNetV3-large) when testing on FRGCv2, and 37.57% and 37.23% respectively when testing on FERET — the highest EERs in both schemes. The paper states this means relying only on synthetic data for S-MAD is not the best option in operational scenarios.
• Best result with FERET training / FRGCv2 testing: the lowest D-EER was 6.09% with EfficientNet-B2 and 6.10% with MobileNetV3-large, both obtained by adding 75% synthetic non-morphed data (1,200 added samples, 2,787 total bona fide). Mean EER across the experiments was 6.86% for EfficientNet-B2 and 6.93% for MobileNetV3-large.
• Best result with FRGCv2 training / FERET testing: the lowest D-EER was 8.68% with EfficientNet-B2 and 10.20% with MobileNetV3-large, both obtained by adding only 10% synthetic data (300 added samples, 3,252 total bona fide). Mean EER was 10.46% for EfficientNet-B2 and 12.05% for MobileNetV3-large.
• The optimal amount of synthetic data depends on the dataset: 75% was best in one training scheme and 10% in the other, indicating that the size of the synthetic sample added is contingent upon the specific dataset used for training.
• Incremental addition does not improve monotonically: for example, with FERET training and FRGCv2 testing, EfficientNet-B2 went from a D-EER of 6.47% with no synthetic data to 8.20% at 10%, 6.47% at 20%, 7.86% at 30%, 6.09% at 50%, 6.09% at 75% and 6.81% at 100%; MobileNetV3-large degraded to 8.17% at 100%.
• Compact architectures are viable: the authors conclude that architectures such as EfficientNet and MobileNet provide strong baselines for morphing attack detection while remaining suitable for realistic deployment scenarios.

Methodology in Plain English

The researchers used three datasets: FERET (529 subjects) and FRGCv2 (533 subjects), both with morphed variants produced by four morphing tools, and the fully synthetic SMDD dataset. Each real dataset was processed in three ways — print/scan at 300 dpi, print/scan at 600 dpi and digital resize — yielding 1,587 bona fide and 6,348 morph images for FERET and 2,952 bona fide and 11,568 morph images for FRGCv2, for overall totals of 19,539 bona fide and 42,916 morph images across the datasets used.

Training ran in two cross-dataset rounds: train on FERET and test on FRGCv2, then the reverse. Images were aligned with MTCNN using a scaling factor of 0.9 to produce 369×369 crops, then augmented (resized to 224×224, random horizontal flip, random rotation, colour jitter, and normalization with ImageNet-1K statistics). An 80/20 split of the training set produced a validation set. Two CNNs were fine-tuned from ImageNet-1K weights using the Adam optimizer (β1 = .99, β2 = .999), categorical cross-entropy loss, batch size 64, 100 epochs, and a learning rate grid of 1e-3, 1e-4 and 1e-5, with 1e-5 giving the best results; Stochastic Gradient Descent was also tested. Runs used an NVIDIA A100-80GB GPU with 8 workers and took approximately seven days in total.

Three experiment scenarios were run: no synthetic data, incremental addition of random samples from the SMDD non-morphed subset (10%, 20%, 30%, 50%, 75%, 100%), and training only on synthetic data. Performance was reported as MACER, BPCER (per ISO/IEC DIS 20059), BPCER at operating points 5, 10 and 20, DET curves, and the D-EER where MACER and BPCER are equal.

Why This Matters

• Research impact: the work quantifies a practical alternative to privacy-restricted bona fide data and shows that indiscriminate use of synthetic data can lead to suboptimal performance, while a carefully controlled amount can improve generalization — a result relevant to how synthetic datasets such as SMDD and the SYN-MAD competition resources are used.
• Border control and immigration checks: morphing is directly relevant to passport and visa document fraud, where an automated S-MAD check must run under time and hardware constraints.
• Mobile identity verification: the paper explicitly targets scenarios such as mobile applications and mobile identity verification, where compact models (7.7M and 2.9M parameters) matter.
• Airport and remote onboarding pipelines: identity verification systems that capture a single face image are the exact setting S-MAD addresses.
• Industry relevance: vendors of biometric verification and identity-document issuance tools gain guidance on whether synthetic data can reduce their dependence on sensitive, privacy-regulated genuine face image collections — and evidence that fully synthetic training is not a substitute.

Future Directions

1. Extend the evaluation with additional backbone models to better understand the impact of model capacity.
2. Incorporate augmentation strategies that approximate operational conditions, such as compression and mild print-scan artefacts.
3. Determine how to select the right amount of synthetic data per training dataset, since the best proportion differed sharply between the two cross-dataset schemes (75% versus 10%).
4. Investigate why synthetic-only training performs so poorly and whether better synthetic generation or mixing strategies can close the gap to real-data training.

Target Audience

Researchers and practitioners in biometric security, face morphing attack detection, and presentation attack detection; engineers building identity verification or document issuance systems; and machine learning practitioners interested in the limits of synthetic data as a privacy-preserving substitute for real training data.

Authors’ abstract

This paper investigates the use of synthetic face data to enhance Single-Morphing Attack Detection (S-MAD), addressing the limitations of availability of large-scale datasets of bona fide images due to privacy concerns. Various morphing tools and cross-dataset evaluation schemes were utilized to conduct this study. An incremental testing protocol was implemented to assess the generalization capabilities as more and more synthetic images were added. The results of the experiments show that generalization can be improved by carefully incorporating a controlled number of synthetic images into existing datasets or by gradually adding bona fide images during training. However, indiscriminate use of synthetic data can lead to sub-optimal performance. Evenmore, the use of only synthetic data (morphed and non-morphed images) achieves the highest Equal Error Rate (EER), which means in operational scenarios the best option is not relying only on synthetic data for S-MAD.