Fixing Domain Bias for Generalized Deepfake Detection

Abstract

Generalizing deepfake detection has posed a great challenge to digital media forensics, as inferior performance is obtained when training sets and testing sets are domain-mismatched. In this paper, we show that a CNN-based detection model can significantly improve performance by fixing domain bias. Specifically, we propose a novel Fixing Domain Bias network (FDBN). FDBN does not rely on manual features, but is based on three core designs. Firstly, a domain-invariant network based on randomly stylized normalization is devised to constrain the domain discrepancy in the feature space. Then, through adversarial learning, a generalizing representation in the stylized distribution is learned to enhance the shared feature bias among manipulation methods in the domain-specific network. Finally, to encourage equality of biases among different domains, we utilize the bias extrapolation penalty strategy by suppressing the expected bias on the extremely-performing domains. Extensive experiments demonstrate that our framework achieves effectiveness and generalization towards unseen face forgeries.