Explainable Deepfake Detection Challenge

View on arXiv ← Back to list

Authors: Abhijeet Narang, Kartik Kuckreja, Shreya Ghosh, Muhammad Haris Khan, Usman Tariq, Jianfei Cai, Abhinav Dhall

Published: 2026-07-23 07:42:43+00:00

Comment: 5 pages, 1 figure

AI Summary

The Explainable Deepfake Detection Challenge, built on the XPlainVerse benchmark, aims to evaluate systems that can not only detect deepfakes but also provide natural-language explanations for their decisions. This challenge addresses the need for both complex technical explanations and concise general-user explanations by evaluating classification performance alongside semantic similarity, simplicity, and grounding metrics of the generated explanations. The methodologies developed will contribute to the advancement of next-generation explainable deepfake detectors.

Abstract

Deepfake detection is moving beyond binary classification decisions toward systems that can also explain the visual evidence supporting those decisions. This transition is important for real-world verification settings, where diverse users need to understand not only whether an image is manipulated, but also why it is considered suspicious. The Explainable Deepfake Detection Challenge at ACM Multimedia 2026 is designed to benchmark this joint capability. Built on XPlainVerse, a million-scale benchmark for explainable deepfake detection, the challenge evaluates methods on image classification and grounded natural-language explanation generation. Participants submit a real/fake label together with two explanations for each image: a detailed complex explanation for technical users and a concise simple explanation for general users. The evaluation combines classification metrics with semantic similarity, simplicity, and intent-aware grounding metrics that assess whether explanations identify the relevant manipulated entities and supporting visual evidence. The methodologies developed through the challenge will contribute to the development of next-generation explainable deepfake detectors. Evaluation script, baseline models, and accompanying code are available on https://github.com/Abhijeet8901/XPlainVerse-ACMChallenge.

Key findings

The challenge results indicate that strong deepfake classification performance alone does not guarantee high-quality, grounded explanations. The best-performing systems, like Pixel Sleuth, effectively combine robust detection (often using DINOv3 backbones) with explanation models (like Qwen-family) optimized for both semantic fidelity and grounding. This highlights the importance of developing integrated systems that can both accurately detect manipulations and clearly articulate the visual evidence supporting those decisions for diverse user needs.

Approach

The challenge evaluates deepfake detection systems on two main tasks: image-level authenticity classification and natural-language explanation generation. Participants submit a real/fake label and two explanations (detailed for technical users and simple for general users) for each image. Evaluation combines detection F1-score with metrics assessing explanation quality, including BERTScore for semantic similarity, Simplicity Level Estimate (SLE) for readability, and LLM-based EntityScore and EvidenceScore for grounding.

Datasets

XPlainVerse (a curated subset)

Model(s)

Qwen3-VL-8B, InternVL3.5-14B, DINOv3 (used by top submissions), Qwen3-VL-32B, Qwen3.5-4B

Author countries

Australia, United Arab Emirates

← Previous