Scholarlink

<p>Face recognition systems, while extensively employed in areas such as secure identity authentication and mobile device access, remain vulnerable to various forms of spoofing attacks, including but not limited to print attacks, video replays, and 3D mask-based attacks. Additionally, the emergence of novel, unidentified spoofing techniques underscores the pressing need for models with superior generalization capabilities. This challenge is further exacerbated by the scarcity of spoofing images available for training, which amplifies the issue of data imbalance. To address these vulnerabilities, we propose an innovative face anti-spoofing (FAS) framework that harnesses the power of data augmentation and selective feature enhancement. Our approach incorporates three pivotal advancements: (1) leveraging the P-FAD dataset we curated, which encompasses a diverse array of spoofing techniques, in conjunction with a robust data augmentation pipeline that synthesizes additional samples from the original dataset; (2) introducing an In-training Selective Feature Augmentation (ISFA) mechanism, designed to dynamically expand the feature space during training and thereby improve the model’s learning potential; and (3) implementing the LaSupCon loss function, meticulously crafted to maximize the utility of augmented data and enhance model performance. Moreover, by integrating these strategies with a dual-component feature extraction framework that captures both content and style features, our method demonstrates significant improvements in generalization across various spoofing methods and domains. Extensive experimental results demonstrate the effectiveness of our framework, showing average improvements of 2.19% in HTER, 2.71% in AUC, and 2.47% in F1-score. These results underscore substantial advancements in accuracy and robustness for cross-domain face anti-spoofing scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Novel face anti-spoofing by integration of data augmentation, selective feature learning, and LaSupCon loss

  • 期刊论文

  • 发表日期：

    2025年8月6日

查看全文
  • Bao Bui-Quoc,
  • Khanh Nguyen-Gia,
  • Hoai Nam Vu

摘要

Face recognition systems, while extensively employed in areas such as secure identity authentication and mobile device access, remain vulnerable to various forms of spoofing attacks, including but not limited to print attacks, video replays, and 3D mask-based attacks. Additionally, the emergence of novel, unidentified spoofing techniques underscores the pressing need for models with superior generalization capabilities. This challenge is further exacerbated by the scarcity of spoofing images available for training, which amplifies the issue of data imbalance. To address these vulnerabilities, we propose an innovative face anti-spoofing (FAS) framework that harnesses the power of data augmentation and selective feature enhancement. Our approach incorporates three pivotal advancements: (1) leveraging the P-FAD dataset we curated, which encompasses a diverse array of spoofing techniques, in conjunction with a robust data augmentation pipeline that synthesizes additional samples from the original dataset; (2) introducing an In-training Selective Feature Augmentation (ISFA) mechanism, designed to dynamically expand the feature space during training and thereby improve the model’s learning potential; and (3) implementing the LaSupCon loss function, meticulously crafted to maximize the utility of augmented data and enhance model performance. Moreover, by integrating these strategies with a dual-component feature extraction framework that captures both content and style features, our method demonstrates significant improvements in generalization across various spoofing methods and domains. Extensive experimental results demonstrate the effectiveness of our framework, showing average improvements of 2.19% in HTER, 2.71% in AUC, and 2.47% in F1-score. These results underscore substantial advancements in accuracy and robustness for cross-domain face anti-spoofing scenarios.