Deepfake detectors improve by preserving diverse forgery clues
Revisiting Cross-Reconstruction for Generalizable Deepfake Detection
Computer Vision and Pattern Recognition
Summary
Detecting fake images made by AI can be tricky because different fake image creators leave different kinds of clues. Previous methods tried to find common patterns by forcing these clues to look the same, but this paper shows it’s better to keep those differences. The authors propose a new way that pays attention to the unique details of each fake image and uses them to better spot forgeries. Their approach helps detectors work well even on new types of fake images they've never seen before.
What this means in practice
- •For digital forensics teams: Enhance detection tools to reliably identify manipulated images from diverse and unseen generation methods using preserved artifact diversity.
- •For social media platform security teams: Improve automated filters to catch AI-generated fake images even if the fakes come from new or different fake image generators.
Authors
Bingjian Yang, Shilei Zhao, Zheng Wang
Abstract
Existing image forgery detectors often suffer from generalization to unseen manipulation methods due to the limited ability to capture transferable forensic cues. Recent cross-reconstruction based methods attempt to improve generalization through semantic-artifact disentanglement, but typically align heterogeneous artifacts across generators and exclude artifact representations during reconstruction, which may overlook the inherent diversity and visual cues of manipulation artifacts. In this work, we revisit cross-reconstruction and introduce an artifact-oriented disentanglement framework for robust image forgery detection. We argue that \textbf{artifact diversity}, i.e., the intrinsic variations of manipulation artifacts introduced by different generation processes, contains complementary forensic cues rather than undesirable domain variations. Instead of enforcing explicit artifact alignment, our framework preserves diverse artifact characteristics through semantically aligned cross-generator reconstruction. Furthermore, we incorporate artifact representations into the reconstruction process and introduce a masked frequency-aware reconstruction strategy to emphasize manipulation-related residuals while reducing semantic interference. This design enables the model to learn transferable forensic representations from diverse artifacts. Extensive experiments on multiple benchmark datasets demonstrate improvements under both cross-dataset and cross-generator evaluation settings. Further analysis and ablation studies validate the effectiveness of artifact diversity preservation and artifact-aware cross-reconstruction.