Revelio: a Modular and Effective Framework for Reproducible Training and Evaluation of Morphing Attack Detectors

Morphing Attack, i.e. the possibility of eluding face verification systems through a facial morphing operation between a criminal and an accomplice, has recently emerged as a serious security threat. Despite the importance of this kind of attack, the development and comparison of Morphing Attack Detection (MAD) methods is still an arduous task, mainly due to the scarcity of publicly-available datasets and the failure of the internal ones to accurately reflect the problem's complexity; these two causes combined lead to low generalization capabilities and challenges in comparing the different MAD approaches proposed in the literature. Therefore, in this thesis, we propose and publicly release Revelio, a flexible and modular framework for the reproducible development and evaluation of both single-image (S-MAD) and differential (D-MAD) systems. Then, we conduct a review of the datasets exploited in the literature and introduce two new ones, namely ChiMo and FEI. Moreover, we introduce a new metric useful for summarizing and simplifying the comparison of diverse approaches across different datasets, named Weighted Average Error across Datasets (WAED), and conduct a review of the publicly available benchmarks used to test algorithms for this task. Besides, an extensive analysis of several state-of-the-art approaches through Revelio is performed, comparing several literature methods and thus deeply analyzing the main challenges in the MAD task. Finally, by exploiting Revelio features, a new model is proposed to improve the state of the art on SOTAMD single-image and double-image benchmarks.