Dealing with Subject Similarity in Differential Morphing Attack Detection

TL;DR

Dealing with Subject Similarity in Differential Morphing Attack Detection tackles robustness of Differential MAD (D-MAD) under high identity similarity, particularly when document images are compared against both criminals and accomplices. The authors introduce ACIdA, a modular system with three components—Attempt Classification (AC), Identity-Artifact (IdA), and Identity (Id)—that are fused via a weighted strategy to robustly detect morphing. Extensive cross-dataset experiments demonstrate state-of-the-art performance on FEI Morph and reveal the added value of combining artifact cues with identity features, especially for the challenging accomplice scenario. The work broadens the practical applicability of D-MAD to enrollment and accomplice detection, and highlights avenues for improving attempt-type classification and deploying the approach on real platforms.

Abstract

The advent of morphing attacks has posed significant security concerns for automated Face Recognition systems, raising the pressing need for robust and effective Morphing Attack Detection (MAD) methods able to effectively address this issue. In this paper, we focus on Differential MAD (D-MAD), where a trusted live capture, usually representing the criminal, is compared with the document image to classify it as morphed or bona fide. We show these approaches based on identity features are effective when the morphed image and the live one are sufficiently diverse; unfortunately, the effectiveness is significantly reduced when the same approaches are applied to look-alike subjects or in all those cases when the similarity between the two compared images is high (e.g. comparison between the morphed image and the accomplice). Therefore, in this paper, we propose ACIdA, a modular D-MAD system, consisting of a module for the attempt type classification, and two modules for the identity and artifacts analysis on input images. Successfully addressing this task would allow broadening the D-MAD applications including, for instance, the document enrollment stage, which currently relies entirely on human evaluation, thus limiting the possibility of releasing ID documents with manipulated images, as well as the automated gates to detect both accomplices and criminals. An extensive cross-dataset experimental evaluation conducted on the introduced scenario shows that ACIdA achieves state-of-the-art results, outperforming literature competitors, while maintaining good performance in traditional D-MAD benchmarks.

Figures (6)

• Figure 1: In traditional D-MAD benchmarks, usually the document image is compared only with bona fide and criminal live photos. Therefore, in this paper, we introduce a new challenging scenario in which the document image is compared with both the criminal and the accomplice. From a practical point of view, this scenario broadens the applications of D-MAD systems.
• Figure 2: Performance of literature D-MAD methods across different identity similarity scores (x-axis) between input subjects. Values close to $1$ indicate a high similarity. MAD performance is expressed as the weighted average of the error-based metrics commonly used in the MAD task, and here referred to as WAE (see Sect. \ref{['sec:metrics']}). As reported, all methods are negatively influenced by the increasing identity similarity, which corresponds to the increase in the accomplice's presence in input images instead of the criminal. The criminal-accomplice ratio is computed as the number of pairs with the criminal and the accomplice over the total. Both percentages are represented with dotted lines. The analyzed D-MAD methods are as follows: ArcFace scherhag2020deep (circle), MagFace kessler2023towards (triangle), Demorphing ferrara2017face (rhombus), and Siamese borghi2021double (square).
• Figure 3: Sample of morphed images that successfully fool the identity-based D-MAD system scherhag2020deep but that exhibit visible artifacts related to the morphing procedure. Hence, the intuition to exploit also artifact detection techniques to improve the performance in the D-MAD task, as described in Section \ref{['sec:mad_issues']}.
• Figure 4: General overview of the proposed system, named ACIdA. As shown, the method is composed of three different modules: the Attempt Classification (AC) module, responsible to determine if the document image is compared with the criminal or the accomplice trust live image (see Sect. \ref{['sec:ac_module']}); the Identity (Id) module, an identity comparison-based MAD system (see Sect. \ref{['sec:id_module']}) and the Identity-Artifact (IdA) module, that integrates both information about identity and artifact detection (see Sect. \ref{['sec:Ida_module']}). Finally, the score of these two MAD modules is combined through a weighted sum to produce the final output of the system, i.e. predicting if the input document image is morphed or bona fide.
• Figure 5: In a D-MAD approach, the morphed image (Fig. \ref{['fig:morphed']}) can be compared with the criminal (Fig. \ref{['fig:criminal']}) or with the accomplice (Fig. \ref{['fig:acc']}), or both, as in the introduced scenario.