| 引用本文： |
-
蔡体健,陈均,罗词勇,刘遵雄,陈子涵.引入全局语义增强的人脸欺诈特征提取研究[J].信息安全学报,2025,10(2):127-138 [点击复制]
- CAI Tijian,CHEN Jun,LUO Ciyong,LIU Zunxiong,CHEN Zihan.Research on Introducing Global Semantic Enhancement for Face Fraud Feature Extraction[J].Journal of Cyber Security,2025,10(2):127-138 [点击复制]
|
|
| 摘要: |
| 基于人脸反欺诈的领域知识,针对人脸活体检测中特征在网络中逐层稀释的问题,该文提出了基于语义增强和交叉注意力优化的人脸活体检测模型。具体来说,首先利用活体样本无欺诈噪声的先验知识,采用活体人脸的半边约束方法来提取欺诈强相关特征;利用欺诈特征的全局移不变性特点,结合深度度量学习技术以及异常检测等方法,该文在U-Net瓶颈层添加语义增强模块来增强欺诈特征,捕获长距离的移不变性特征,同时对比了三个不同的语言增强模块在模型上的性能,然后在编码块和解码块之间的跳跃连接后引入交叉自注意力模块,以进一步增强全局的欺诈信息和重要区域的关注。此外,该文将U-Net模型的解码块中的传统卷积算子替换为中心差分卷积算子,以提取细粒度的欺诈特征,并通过计算中心像素与周围像素之间的差异,去除光照、环境的影响,以此提高模型的鲁棒性能。经过在四个常用的人脸活体检测数据集CASIA-MFSD、MSU-MFSD、OULU-NPU、Replay-Attack上测试与评估,进行了数据集内实验、跨数据集实验和消融实验等,对模型进行了复杂度分析以及对部分实验进行了可视化分析,该文模型能够有效降低人脸分类的错误率。 |
| 关键词: 人脸活体检测 全局语义增强 交叉注意力 中心差分卷积 深度度量学习 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2025.03.09 |
| 投稿时间：2023-07-11修订日期：2023-09-23 |
| 基金项目:本课题得到国家自然科学研究基金(No. 62162026), 江西省自然科学基金资助项目(No. 20232BAB202055, No. 20242BAB25114)资助。 |
|
| Research on Introducing Global Semantic Enhancement for Face Fraud Feature Extraction |
|
CAI Tijian, CHEN Jun, LUO Ciyong, LIU Zunxiong, CHEN Zihan
|
| (School of Information Engineering, East China Jiao Tong University, Nanchang 330013, China) |
| Abstract: |
| Based on the domain knowledge of face anti-fraud, this paper proposes a face liveness detection model based on semantic enhancement and cross-attention optimization to address the problem of feature dilution layer by layer in the network during face liveness detection. Specifically, we first use the prior knowledge that living samples are free of fraud noise, and use the half-edge constraint method of living faces to extract strong correlation features of fraud; we also use the global shift invariance characteristics of fraud features, combined with deep metric learning technology and anomaly detection, etc. Method, this article adds a semantic enhancement module to the U-Net bottleneck layer to enhance fraud features and capture long-distance shift invariance features. At the same time, it compares the performance of three different language enhancement modules on the model, and then in the encoding block and decoding A cross-self-attention module is introduced after the skip connection between blocks to further enhance the global fraud information and focus on important areas. In addition, this paper replaces the traditional convolution operator in the decoding block of the U-Net model with a central difference convolution operator to extract fine-grained fraud features and remove them by calculating the difference between the central pixel and the surrounding pixels. The influence of lighting and environment is used to improve the robust performance of the model. After testing and evaluation on four commonly used face liveness detection data sets CASIA-MFSD, MSU-MFSD, OULU-NPU, and Replay-Attack, intra-dataset experiments, cross-dataset experiments, and ablation experiments were conducted to verify the model. Complexity analysis and visual analysis of some experiments were conducted. The model in this article can effectively reduce the error rate of face classification. |
| Key words: face liveness detection global semantic enhancement cross-attention central difference convolution deep metric learning |
