V-MAD: Video-based Morphing Attack Detection in Operational Scenarios

TL;DR

The paper introduces Video-based Morphing Attack Detection (V-MAD) to exploit sequences of frames in operational settings like border control gates, showing that multi-frame information improves robustness to pose and illumination variability. It adapts existing Differential MAD (D-MAD) models to the V-MAD task by computing per-frame scores and fusing them through simple or quality-aware strategies, and explores machine learning fusion with an SVM to further enhance performance. A new real-world operational database is collected to evaluate V-MAD under realistic conditions, including various face image quality metrics (MagFace, CR-FIQA, SER-FIQ) and ISO components (illumination, defocus, pose). Key findings indicate that average/median frame fusion often surpasses single-frame D-MAD, quality information yields additional gains, and ML-based fusion can outperform traditional fusion, though an oracle-like bound remains distant, underscoring the need for more robust V-MAD methods and datasets. Overall, V-MAD represents a significant advancement toward robust morphing attack detection in real-world video acquisition scenarios, with practical implications for ABC gate security and biometric verification systems.

Abstract

In response to the rising threat of the face morphing attack, this paper introduces and explores the potential of Video-based Morphing Attack Detection (V-MAD) systems in real-world operational scenarios. While current morphing attack detection methods primarily focus on a single or a pair of images, V-MAD is based on video sequences, exploiting the video streams often acquired by face verification tools available, for instance, at airport gates. Through this study, we show for the first time the advantages that the availability of multiple probe frames can bring to the morphing attack detection task, especially in scenarios where the quality of probe images is varied and might be affected, for instance, by pose or illumination variations. Experimental results on a real operational database demonstrate that video sequences represent valuable information for increasing the robustness and performance of morphing attack detection systems.

Figures (5)

• Figure 1: In operational scenarios, V-MAD represents a viable paradigm when a single probe document image is compared with an input sequence to detect whether the image is morphed or not. V-MAD differs from currently available literature solutions, based only on a single (S-MAD) or a pair (D-MAD) of images.
• Figure 2: Practical implementation of the V-MAD task. Each frame of the input sequence is analyzed by the same D-MAD model, which receives also the document image as additional input and by a quality tool. Both output MAD and quality scores are then combined through a specific fusion strategy to produce in output the final single score.
• Figure 3: An example of the quality scores computed on a sequence of frames. As reported in Section \ref{['sec:quality']}, the first three methods (CRFIQA CR-FIQA, MagFace MagFace and SERFIQ SER-FIQ) are able to compute an overall quality score on the whole image, while the last three are based on specific aspects of the image.
• Figure 4: V-MAD performance comparison of three D-MAD models using different MAD score fusion strategies (see Sect. \ref{['sec:score_fusion']}). The dashed line represents the theoretical upper bound of performance, while the dotted line, based on a random choice in the score set, represents the lower bound. Metrics are expressed as errors, then lower are better.
• Figure 5: DET curves for the different V-MAD approaches obtained by exploiting image quality estimated as either unified (first row) or specific (second row) quality measures.