A carregar...
Publicação
Face morphing detection in identification documents
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 31.34 MB | Adobe PDF |
Autores
Resumo(s)
Os ataques de morphing facial constituem uma ameaça crítica aos sistemas de autenticação biométrica, permitindo a criação de imagens compostas que correspondem a múltiplas identidades. Este trabalho investiga a deteção de ataques de morphing (MAD) baseada em deep learning, com foco nas arquiteturas Vision Transformer (ViT) e na interpretabilidade através de heatmaps baseados em oclusões.
Uma revisão sistemática de 23 estudos evidenciou limitações recorrentes: reduzida disponibilidade de código, análise de interpretabilidade insuficiente e dificuldades na generalização entre diferentes datasets. Estes resultados motivaram cinco experiências que avaliaram estratégias de pré-processamento, composição de dados, equilíbrio de classes e robustez entre datasets, incluindo uma experiência com um extrator ViT seguido de um classificador Support Vector Machine (SVM).
Os resultados demonstram que o isolamento da face melhora consistentemente o desempenho, e que a diversidade dos dados aumenta a generalização. Embora o modelo ViT tenha alcançado precisão quase perfeita em avaliações intra-dataset, o desempenho degradou-se substancialmente em conjuntos de dados não incluídos no treino, evidenciando o impacto da mudança de domínio. A inclusão de técnicas avançadas de morphing durante o treino melhorou a deteção de ataques mais simples, enquanto que a sua não inclusão proporcionou uma generalização limitada.
Para responder aos requisitos de interpretabilidade em contextos críticos, este trabalho propõe uma plataforma de análise padronizada baseada em oclusões, combinando estratégias semânticas e espaciais. Aplicados ao modelo ViT-SVM, os heatmaps indicam que as regiões do nariz e da boca são especialmente relevantes, com o modelo a focar-se em áreas de transição de profundidade. A plataforma é independente do modelo e reproduzível, oferecendo uma abordagem sistemática para identificar regiões faciais que impactam o desempenho do classificador.
Em suma, este trabalho valida arquiteturas ViT para MAD, analisa os fatores chave que influenciam a robustez dos modelos e desenvolve ferramentas de interpretabilidade para implementar sistemas de segurança biométrica fiáveis.
Face morphing attacks pose a critical security threat to biometric authentication systems by enabling the creation of composite facial images that can be matched to multiple identities. This work investigates deep learning-based Morphing Attack Detection (MAD), with a focus on Vision Transformer (ViT) architectures and on improving model interpretability through occlusion-based heatmap analysis. A systematic review of 23 recent studies highlighted several persistent research limitations: scarce code availability hindering reproducibility, insufficient interpretability analysis, and ongoing challenges in cross-dataset generalization. These findings informed five experiments evaluating preprocessing strategies, dataset composition, class balancing, and cross-dataset robustness, with one experiment specifically employing a ViT feature extractor coupled with a Support Vector Machine (SVM) classifier. The results show that face cropping consistently improves detection performance, and that dataset diversity enhances generalization capability. While the ViT-based model achieved near-perfect accuracy in intra-dataset evaluations, performance degraded substantially on unseen datasets, underscoring the impact of domain shift. Additionally, training on advanced morphing techniques improved detection of simpler ones, whereas training on more simple techniques offered limited transferability. To address interpretability requirements in security-critical contexts, this work proposes a standardized occlusion-based analysis framework that combines landmarkbased (semantic) and grid-based (spatial) strategies. When applied to the ViT-SVM model, the resulting heatmaps indicate that the nose and mouth regions are especially influential for detection decisions, with the model focusing on depth-transition areas. The framework is model-agnostic and reproducible, offering a clear and systematic approach to identifying facial regions that contribute to MAD classifications. Overall, this work advances the field by validating ViT architectures for MAD, analyzing key factors affecting detection robustness, and developing interpretability tools essential for the deployment of trustworthy deep learning-based biometric security systems.
Face morphing attacks pose a critical security threat to biometric authentication systems by enabling the creation of composite facial images that can be matched to multiple identities. This work investigates deep learning-based Morphing Attack Detection (MAD), with a focus on Vision Transformer (ViT) architectures and on improving model interpretability through occlusion-based heatmap analysis. A systematic review of 23 recent studies highlighted several persistent research limitations: scarce code availability hindering reproducibility, insufficient interpretability analysis, and ongoing challenges in cross-dataset generalization. These findings informed five experiments evaluating preprocessing strategies, dataset composition, class balancing, and cross-dataset robustness, with one experiment specifically employing a ViT feature extractor coupled with a Support Vector Machine (SVM) classifier. The results show that face cropping consistently improves detection performance, and that dataset diversity enhances generalization capability. While the ViT-based model achieved near-perfect accuracy in intra-dataset evaluations, performance degraded substantially on unseen datasets, underscoring the impact of domain shift. Additionally, training on advanced morphing techniques improved detection of simpler ones, whereas training on more simple techniques offered limited transferability. To address interpretability requirements in security-critical contexts, this work proposes a standardized occlusion-based analysis framework that combines landmarkbased (semantic) and grid-based (spatial) strategies. When applied to the ViT-SVM model, the resulting heatmaps indicate that the nose and mouth regions are especially influential for detection decisions, with the model focusing on depth-transition areas. The framework is model-agnostic and reproducible, offering a clear and systematic approach to identifying facial regions that contribute to MAD classifications. Overall, this work advances the field by validating ViT architectures for MAD, analyzing key factors affecting detection robustness, and developing interpretability tools essential for the deployment of trustworthy deep learning-based biometric security systems.
Descrição
Palavras-chave
Deteção de ataques de morphing Morphing facial Deep learning Interpretabilidade Heatmaps
Contexto Educativo
Citação
Editora
Licença CC
Sem licença CC