FD-MAD: Frequency-Domain Residual Analysis for Face Morphing Attack Detection

TL;DR

This paper tackles the cross-dataset generalization problem in single-image face morphing attack detection (S-MAD) by introducing a frequency-domain residual framework. It computes global radial Fourier residuals after removing a power-law baseline and augments them with region-wise spectral descriptors for four facial regions, fused via a Markov Random Field to enforce spatial consistency. The approach, trained on SMDD and evaluated on FRLL-Morph and MAD22 (including MorDIFF), achieves a best-average EER of $1.85\%$ on FRLL-Morph and $6.12\%$ on MAD22, while maintaining favorable BPCER at fixed APCER, demonstrating strong cross-morph generalization with a lightweight model. The findings indicate that Fourier-domain residual modeling, combined with structured regional fusion, offers a competitive and interpretable alternative to deep architectures in resource-constrained S-MAD deployments, with potential extensions to 3D cues and contrastive learning.

Abstract

Face morphing attacks present a significant threat to face recognition systems used in electronic identity enrolment and border control, particularly in single-image morphing attack detection (S-MAD) scenarios where no trusted reference is available. In spite of the vast amount of research on this problem, morph detection systems struggle in cross-dataset scenarios. To address this problem, we introduce a region-aware frequency-based morph detection strategy that drastically improves over strong baseline methods in challenging cross-dataset and cross-morph settings using a lightweight approach. Having observed the separability of bona fide and morph samples in the frequency domain of different facial parts, our approach 1) introduces the concept of residual frequency domain, where the frequency of the signal is decoupled from the natural spectral decay to easily discriminate between morph and bona fide data; 2) additionally, we reason in a global and local manner by combining the evidence from different facial regions in a Markov Random Field, which infers a globally consistent decision. The proposed method, trained exclusively on the synthetic morphing attack detection development dataset (SMDD), is evaluated in challenging cross-dataset and cross-morph settings on FRLL-Morph and MAD22 sets. Our approach achieves an average equal error rate (EER) of 1.85\% on FRLL-Morph and ranks second on MAD22 with an average EER of 6.12\%, while also obtaining a good bona fide presentation classification error rate (BPCER) at a low attack presentation classification error rate (APCER) using only spectral features. These findings indicate that Fourier-domain residual modeling with structured regional fusion offers a competitive alternative to deep S-MAD architectures.

Figures (5)

• Figure 1: Comparison between frequency-domain data of a bona fide face and a morph sample. The key observation is that the morphs typically differ from the bona fide samples in the mid-high frequencies, which is used as the insight for the proposed method.
• Figure 2: Overview of the proposed Fourier‐domain morphing detector. The input face is transformed to the frequency domain and summarized into (top) global azimuthally averaged FFT magnitude profiles used by an SVM classifier to obtain a global score ($s_{global}$), and (bottom) local profiles extracted from the left eye, right eye, nose, and mouth, whose region-wise classifiers are fused by a Markov Random Field (MRF) to produce a local score ($s_{local}$). The global and local scores are then combined through a weighted fusion to obtain the final morphing detection score.
• Figure 3: Residual radial-frequency profiles of a bona fide image and its corresponding morph. Compared to the bona fide image, the morph exhibits systematic deviations, particularly in the mid-to-high frequency bands, consistent with blending and interpolation artifacts introduced during morphing.
• Figure 4: Exact MRF inference time versus number of facial regions $R$. While the complexity scales exponentially with $R$ due to full enumeration, faces admit only a limited number of meaningful regions, and inference remains tractable even for conservative upper bounds on $R$.
• Figure 5: t-SNE visualization of frequency-domain global features obtained from different morphing attacks of FRLL-Morph frll_morph (left) and MAD22 synmad22 (right). For MAD22, the WebMorph attack follows a similar distribution, making it a hard one.