Role of Artifact Representations in Reconstruction

Determine whether forgery-artifact representations should participate in the reconstruction process for image forgery detection, rather than being excluded as reconstruction-irrelevant information.

Background

Cross-reconstruction methods for deepfake detection commonly disentangle semantic information from manipulation artifacts and often exclude artifact representations from the reconstruction process. The paper argues that this practice may discard meaningful visual forensic cues, including color mismatches and detail distortions, that can help identify manipulated images. The unresolved issue is whether incorporating artifact representations into reconstruction improves the learning of transferable forensic representations and supports more robust detection of unseen manipulations.