MSU Team at the Explainable Deepfake Detection Challenge 2026: Grounded Artifact Evidence for Deepfake Detection

Authors: Artem Filippov, Aleksandr Gushchin, Kirill Koltsov, Dmitriy Vatolin, Anastasia Antsiferova

Published: 2026-10-07 12:30:04+00:00

AI Summary

This paper presents a solution for the Explainable Deepfake Detection Challenge, focusing on detecting manipulated images and providing grounded visual explanations. The method employs a multi-backbone detector combining DINOv3 models with Mesorch features, enhanced by Grounding-DINO-based pseudo-mask generation and a patch-level contrastive objective. For explanations, class-conditional Qwen3-VL models are used, followed by a GRPO-optimized text simplification model.

Abstract

Recent advances in generative image models have made many manipulated images highly realistic, raising the need for detectors that are not only accurate but also able to provide visual evidence for their decisions. In this paper, we present our solution to the Explainable Deepfake Detection Challenge [2] on the XPlainVerse dataset [1], where systems are required to predict whether an image is real or fake and generate both complex and simple explanations grounded in visible forensic cues. Our method follows a modular detection-and-explanation design. For the real/fake decision, we build a multi-backbone detector that combines several DINOv3 models with Mesorch manipulation-localization features, bringing together pretrained visual representations, DCT-aware cues, and multi-scale forensic information. To inject explanation evidence into the detector, we use a Grounding-DINO-based pseudo-mask generation pipeline that converts local artifact descriptions from training explanations into weak patch- level supervision for an Artifact Evidence Map. We further introduce a local patch-level contrastive objective that separates artifact and authenticity evidence in the detector feature space without requiring paired images or pixel-level manipulation masks. For language output, we use class-conditional Qwen3-VL models to generate complex explanations for fake and real predictions, followed by a GRPO-optimized text simplification model. The proposed methods were trained and evaluated on the challenge subset of XPlainVerse. On the full test split, our submission achieves 0.9349 detection accuracy, 0.5571 explanation score, and a 0.7456 final challenge score.

Key findings

The proposed system achieved a detection accuracy of 0.9349 and a final challenge score of 0.7456 on the XPlainVerse test set. Ablation studies showed that integrating the Artifact Evidence Map improved AUC from 0.8975 to 0.9139, and the local contrastive loss further boosted it to 0.9417, demonstrating the effectiveness of explanation-grounded supervision and patch-level contrastive learning.

Approach

The approach uses a modular design for detection and explanation. For deepfake detection, a multi-backbone detector integrates DINOv3 models and Mesorch manipulation-localization features. Explainability is achieved by generating pseudo-masks from artifact descriptions to provide weak patch-level supervision and by employing a local patch-level contrastive objective. Complex and simple explanations are generated using class-conditional Qwen3-VL models and a GRPO-optimized text simplifier.

Datasets

XPlainVerse

Model(s)

DINOv3, Mesorch, Grounding DINO, Qwen3-VL (32B-Instruct, 8B-Instruct), GRPO

Author countries

Russia

← Previous