EviStress: Evidence Stress Learning for Generalizable Face Forgery Detection with Textual Explanations

Key points

• The rapid spread of face forgeries on the Web challenges the credibility of visual content.1
• In practical Web environments, manipulated media originate from evolving generation pipelines and undergo diverse processing during dissemination, making reliable authenticity assessment under unseen manipulations and visual changes a critical challenge.1
• We observe that a standard LoRA-adapted CLIP detector exhibits uneven patch-level decision dependence, concen trating a disproportionate share of class-specific relevance on a small fraction of visual patches.1
• We further introduce Discriminative Variation Preservation (DVP), which constrains the residual between clean and perturbed representations to retain authenticity discriminative information.1

Sources (1)

• [1]EviStress: Evidence Stress Learning for Generalizable Face Forgery Detection with Textual Explanations
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 02:41 PM

  To improve robustness to variations in such highly dependent evidence, we propose EviStress, a two-stage framework for generalizable face forgery detection with textual explanations.

  The rapid spread of face forgeries on the Web challenges the credibility of visual content.

Extractive summary: sentences quoted from the sources.