全局视野多层次特征增强的人脸伪造检测方法

Face Forgery Detection Based on Global Viewand Multi-level Feature Enhancement

针对现有的深度伪造检测方法的偏重于局部伪造纹理信息以及对于未知伪造类型人脸泛化性检测精度低的问题,提出了一种基于全局视野的多层次检测网络,利用多头注意力机制聚合空域中Query的像素级别强度以及梯度信息生成Key和Value,使得网络在空域中构建长距离依赖关系便于获取全局伪造信息,并结合多层次特征增强策略对检测网络不同层次之间的提取伪造特征进行增强,用于提升网络的空间感知局部伪造信息能力。实验结果表明在数据集内和跨伪造类型数据集上均有较高的ACC和AUC测试得分,消融研究验证了模型各个子模块的有效性。

The paper proposes a new method for detecting deepfakes that addresses the problems of existing methods,which tend to focus on local forgery texture information and have low generalization accuracy for detecting unknown types of forgeries in faces.The proposed method is based on a multi-level detection network with a global perspective,which uses a multi-head attention mechanism to aggregate pixel-level intensity and gradient information of Query in the spatial domain to generate Key and Value,enabling the network to build long-range dependency relationships in the spatial domain to obtain global forgery information.The multi-level feature enhancement strategy is combined with the detection network to enhance the extraction of forgery features between different levels of the network,improving the network's ability to perceive local forgery information in space.Experimental results show that the proposed method has high ACC and AUC test scores both within the dataset and across different types of forgery datasets,and ablation studies verify the effectiveness of the model's various sub-modules.