Explainable Visual Forgery Detection: A Survey

Abstract: The rapid growth of AI-driven image manipulation technologies poses critical challenges for verifying content authenticity. While many forgery detection systems achieve high accuracy, their black-box nature limits deployment in high-stakes domains that demand transparency and explainability. This survey presents the first comprehensive review of explainable forgery detection in images and videos, introducing a novel taxonomy structured around three dimensions: Forgery Localization (FL), which pinpoints manipulated regions; Forgery Attribution (FA), which identifies manipulation sources; and Forgery Judgment Basis (FJB), which clarifies decision reasoning. We systematically analyze 48 state-of-the-art methods across single-modal and multi-modal settings, examining architectural innovations and explainability mechanisms. Four feature-driven strategies (RGB, frequency-domain, noise-texture, and representation learning) are reviewed in detail, highlighting their complementary strengths. Benchmark datasets and evaluation protocols are also compared, and open challenges are identified, including the need for standardized explanation formats, uncertainty quantification, and broader dataset coverage. By establishing this taxonomy and synthesizing recent progress, this survey lays a foundation for developing transparent and trustworthy forgery detection systems, supporting real-world applications in forensic analysis, news verification, and regulatory compliance.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Ming-Hsuan_Yang1

Submission Number: 7354