Improving Generalization of Deepfake Detectors by Imposing Gradient Regularization

一、研究背景
以往工作通过改进网络架构或者合成新数据来提升伪造检测的泛化性，但是前者需要承担额外的计算开销，后者随着生成技术的发展会逐渐失效。
the previous methods always require network architecture designs or data synthesizing strategies to improve the generalization ability of deepfake detection. The former easily suffer from extra computational overhead in the inference phase, while the latter may be invalid with the advancement of generative methods.

二、研究动机

1.图像纹理可以用实例级的浅层特征的通道均值和方差来表示。
image textures can be represented by instance-level channel-wise means and variances