XPlainVerse: A Million-Scale Benchmark for Explainable Deepfake Detection

2026-07-03

AI Summary

This work addresses the critical gap in existing deepfake detection methods, whose natural language explanations often lack alignment with actual visual artifacts in images and thus suffer from low credibility. While current benchmarks focus solely on classification accuracy, they neglect the faithfulness of explanations. To bridge this gap, we introduce the first million-scale benchmark for explainable deepfake detection, comprising image pairs from five real sources and twelve generative models. We propose an Edit-Check multi-stage editing verification pipeline to ensure tampering validity and define two novel metrics—EntityScore and EvidenceScore—to quantitatively assess the reasoning fidelity of explanations. The benchmark provides both technical and simplified dual-modality explanations and, validated on 2,000 human-annotated samples, demonstrates high consistency with human judgments in explanation quality, thereby establishing a foundation for trustworthy and scalable explainable deepfake detection research.

Abstract

As deepfake detection models increasingly produce natural language explanations, their reasoning often remains weakly grounded in visual artifacts, limiting reliability and user trust. Existing benchmarks mainly evaluate classification accuracy, overlooking whether explanations reflect the actual manipulations. This gap hinders progress toward deployable, explainable deepfake detection systems. To this end, we introduce XPlainVerse, a large-scale benchmark designed for joint deepfake detection and human-centered explanation. XPlainVerse comprises one million real and manipulated images, pairing authentic images from five established sources with forgeries generated by twelve off-the-shelf image editing and synthesis models. We further propose a multi-stage filtering pipeline, Edit-Check, to verify if manipulations satisfy their intended edits, enabling reliable reasoning supervision at scale. Beyond dataset scale, XPlainVerse provides two complementary explanation styles: technical explanations for expert analysis and simplified explanations optimized for non-technical users. To evaluate explanation quality beyond surface similarity, we propose novel metrics, EntityScore and EvidenceScore, that measure reasoning fidelity by checking whether explanations correctly identify manipulated entities and visual evidence. Human annotations on 2,000 explanation pairs validate our dataset quality against human judgment. We believe XPlainVerse will establish grounded explanation quality as a measurable dimension of deepfake detection and support scalable research on trustworthy, interpretable models.

Problem

• explainable deepfake detection
• visual artifacts
• explanation grounding
• manipulation reasoning
• trustworthy AI

Innovation

• explainable deepfake detection
• XPlainVerse
• reasoning fidelity
• Edit-Check pipeline
• EntityScore