Основні поняття
A modular deep learning-based approach, named ACIdA, that effectively detects morphed faces by combining identity features and artifact analysis, addressing the challenge of high subject similarity in differential morphing attack detection scenarios.
Анотація
The paper proposes a novel modular approach, named ACIdA, for Differential Morphing Attack Detection (D-MAD). The key insights are:
-
Traditional D-MAD methods based solely on identity comparison suffer from reduced effectiveness when the morphed image and the live acquisition are highly similar (e.g., when the morphed image is more similar to the accomplice).
-
The proposed ACIdA method combines identity features and artifact analysis to improve robustness against high subject similarity. It consists of three modules:
- Attempt Classification (AC) module: Classifies the input pair as criminal, accomplice, or bona fide attempt.
- Identity-Artifact (IdA) module: Integrates identity features and artifact detection to handle accomplice attempts.
- Identity (Id) module: Relies only on identity comparison, effective for criminal and bona fide attempts.
-
The final score is a weighted combination of the outputs from the IdA and Id modules, using the probabilities from the AC module.
-
Extensive experiments on a new challenging scenario, where the morphed image is compared with both the criminal and the accomplice, demonstrate the effectiveness of the proposed ACIdA method, outperforming state-of-the-art D-MAD approaches.
-
The results highlight the importance of the Attempt Classification module in improving overall performance, and the benefits of combining identity features and artifact analysis to address high subject similarity.
Статистика
The paper reports the following key statistics:
The proposed ACIdA method achieves an EER of 0.070, B0.05 of 0.105, and B0.01 of 0.280 on the challenging FEI Morph dataset.
The ablation study shows that the combination of the IdA and Id modules is essential to obtain good performance.
The accuracy of the Attempt Classification (AC) module is 66.9%, indicating room for improvement in this critical component.
Цитати
"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."
"We observe that, unfortunately, most of the D-MAD approaches available in the literature have been developed thinking of a practical application at the verification stage, i.e. when the criminal subject tries to use the passport for instance at ABC gates. Therefore, the large majority of existing benchmarks only consider the comparison between the document image and the criminal or the bona fide subject."