Face forgery detection is an important yet challenging task that aims to distinguish whether a face video has been modified. As various types of face forgery are constantly produced and made available, existing methods usually overfit to the manipulation methods they are trained for, and cannot generalize well to detect the unseen or unknown forgery types. To address this issue, we present a systematic study on a more generalizable solution of face forgery detection, which endows the model an ability to recognize fake videos with unpredictable forgery types. Specifically, we develop a model-agnostic learning approach with a gradient-based meta-train and meta-test procedure to simulate the domain shift from known to unknown forgery types. To further emphasize the relative importance of different available forgery types during training, we propose a self-guided importance sampling strategy, which is integrated with a general video-level classification network. We also build a dataset with a wide range of 10 different forgery types to benchmark the generalization ability of face forgery detection. Extensive experiments on multiple testing protocols of evaluating generalization ability show that our method generalizes significantly better on unknown forgery manipulations.

人脸伪造检测是一项重要但具有挑战性的任务，旨在区分人脸视频是否被修改。由于各种类型的面部伪造不断产生和可用，现有的方法通常会过度拟合它们所训练的操纵方法，并且不能很好地泛化以检测看不见或未知的伪造类型。为了解决这个问题，我们对更通用的面部伪造检测解决方案进行了系统研究，该解决方案赋予模型识别具有不可预测的伪造类型的假视频的能力。具体来说，我们开发了一种与模型无关的学习方法，该方法使用基于梯度的元训练和元测试程序来模拟从已知伪造类型到未知伪造类型的域转移。为了进一步强调训练期间不同可用伪造类型的相对重要性，我们提出了一种自我引导的重要性采样策略，该策略与通用视频级分类网络相结合。我们还构建了一个包含 10 种不同伪造类型的数据集，以衡量人脸伪造检测的泛化能力。对评估泛化能力的多种测试协议进行的大量实验表明，我们的方法在未知伪造操作上的泛化效果明显更好。