Existing deepfake detection methods perform poorly on face forgeries generated by unseen face manipulation algorithms. The generalization ability of previous methods is mainly improved by modeling hand-crafted artifact features. Such properties, on the other hand, impede their further improvement. In this paper, we propose a novel deepfake detection method named Common Artifact Deepfake Detection Model, which aims to learn common artifact features in different face manipulation algorithms. To this end, we find that the main obstacle to learning common artifact features is that models are easily misled by the identity representation feature. We call this phenomenon Implicit Identity Leakage (IIL). Extensive experimental results demonstrate that, by learning the binary classifiers with the guidance of the Artifact Detection Module, our method effectively reduces the influence of IIL and outperforms the state-of-the-art by a large margin, proving that hand-crafted artifact feature detectors are not indispensable when tackling deepfake problems.
翻译:现有深假检测方法在由隐形面部操纵算法生成的面部伪造中表现不佳。 以往方法的普及能力主要通过手工制作的人工工艺特征模型得到改进。 另一方面,这些特性阻碍了这些特性的进一步改进。 在本文中,我们提议了一种新型的深假检测方法,名为常见人工智能深假检测模型,目的是在不同面部操纵算法中学习共同的人工工艺特征。 为此,我们发现学习常见文物特征的主要障碍是模型很容易被身份标识特征特征特征特征特征特征所误导。 我们称之为“隐形身份渗漏”现象( IIL)。 广泛的实验结果表明,通过在人工智能检测模块的指导下学习二进制分类器,我们的方法有效地减少了IIL的影响力,并大大超越了工艺现状。 证明手工艺特征探测器在解决深层问题时并非不可或缺的。