结合多视角学习与一致性表征的人脸伪造检测

现有的人脸伪造检测方法通常在已知域上表现较好,但面临过拟合的风险,在应对未知场景时无法保持良好的检测能力。为解决此问题,提出一种结合多视角学习与一致性表征的人脸伪造检测框架。为捕获更全面的伪造痕迹,将输入图像转换为两种互补视角并采用双流骨干网络进行多视角特征学习。引入一致性度量,以补丁级监督的方式明确约束不同视角输出的局部特征的相似度。为提高检测精度,采用特征分解策略进一步优化伪造特征,减少不相关因素的干扰,并以伪造特征空间的决策作为最终的预测结果。在基准数据集上进行的大量实验表明,所提出的方法优于现有的主流模型,具备良好的跨域泛化能力。

Combining multi-view learning and consistent representation for face forgery detection

Most of the existing face forgery detection methods usually achieve acceptable detection performance on known attacks, but still face the risk of overfitting and fail to maintain good detection capability when dealing with unknown scenes. To solve this problem, an effective face forgery detection framework based on multi-view learning and consistent representation was proposed. To capture more comprehensive forgery traces, the input image was transformed into two complementary views and a dual-stream backbone network was used for multi-view feature learning. The consistency metric was introduced to explicitly constrain the similarity of local features output from different viewpoints in a patch-level supervised manner. To improve the detection accuracy of the model, the feature decomposition strategy further optimized the forgery-relevant feature to reduce the interference of irrelevant factors, and the decision made from the forgery-relevant feature space was used as the final prediction. Extensive experiments on benchmark datasets show that the proposed method outperforms the existing mainstream approaches with good cross-domain generalization capability.