Explainable face forgery detection aims to provide not only a correct judgment but also forensic evidence that is backed by the input image. However, the cues used to train such systems are typically derived from the forgery generation pipeline. A new arXiv paper argues that these pipeline-derived cues are not automatically evidence for a particular image.

The paper introduces the concept of "cue supportability" to address this gap. It proposes a forensic distillation approach that uses supportability to determine which cues are genuinely grounded in the input image, rather than treating every generation-related cue as valid evidence. The abstract does not include experimental details or quantitative results, so the contribution is currently presented at a conceptual level.

The significance of this idea is that it could make face forgery explanations more trustworthy by ensuring that the evidence shown to users is actually present in the image being analyzed. The paper does not claim that all pipeline cues are useless, but it stresses that supportability should be a key criterion when distilling forensic cues for explainable detection.