EDD | Explainable Deepfake Detection with Feature-robust Augmentation and Evidence-grounded Explanation Optimization

Abstract

Accurate detection with transparent explanations.

This paper introduces a novel framework designed to enhance both the detection robustness and the explanation faithfulness of deepfake detection systems. Existing explainable deepfake detection methods often suffer from two critical vulnerabilities: sensitivity to image quality degradations, which causes prediction instability, and a lack of evidence accuracy verification, leading to explanations that omit key evidence or include hallucinated details.

To solve this, our framework jointly addresses these challenges using two main components: Feature-Robust Augmentation, which employs a degradation-aware augmentation pipeline with a mean-teacher architecture and supervised contrastive loss to prevent intra-instance feature drift; and Evidence-Grounded Explanation Optimization, which constructs chosen-rejected explanation pairs and uses Direct Preference Optimization (DPO) to prioritize genuine manipulation evidence over superficial reasoning patterns.

Through these methods, our framework delivers a detector that is not only accurate across diverse quality conditions but also transparent in its decision-making, achieving state-of-the-art overall performance and ranking first on the public leaderboard for the ACM MM 2026 Explainable Deepfake Detection Challenge.

Pain Point ①: Brittleness of Detection Models under Image Quality Degradation

Deepfake detectors are highly sensitive to image quality degradation. When images undergo common distortions such as compression, blur, or noise, the brittleness of detection models leads to unstable predictions and a significant drop in accuracy.

**Detection accuracy on samples with different image quality.**

**Performance change under increasing augmentations.**

Illustration of analysis for detection robustness. (a) Performance drops under quality degradation. (b) Mean-teacher stabilizes learning under diverse augmentations while the baseline suffers from feature drift.

Pain Point ②: Inadequate Evidence-Verification of Existing Explanation Mechanisms

Prevailing approaches predominantly focus on the procedural format of reasoning—mandating multi-step thinking or sufficient response length—while neglecting the factual accuracy and completeness of the evidence presented within the rationale. We evaluate explanations of the baseline Qwen3-VL-8B-Instruct on a 1k-subset of the validation set, observing two prevalent failure modes:

Evidence omission — crucial manipulation traces are missing from the explanation, present in 68.3% of samples.
Irrelevant information intrusion — spurious details are hallucinated, present in 57.7% of samples.

These deficiencies not only mislead users but also erode the trust that explainability aims to establish.

Highlights

🏆 ACM MM 2026 Challenge Winner

🏆

Winner of ACM-MM 2026 Challenge

Ranked first for the ACM MM 2026 Explainable Deepfake Detection Challenge.

🛡️

Feature-Robust Augmentation

Degradation-aware augmentation pipeline with mean-teacher architecture and supervised contrastive learning for robust detection across diverse quality conditions.

💡

Evidence-Grounded Explanation

DPO-based optimization with chosen-rejected explanation pairs to ensure complete and accurate explanations, plus GRPO for concise explanation generation.

Method

Framework Overview

Our framework jointly addresses detection robustness and explanation faithfulness through two main components.

Degradation-Aware Augmentation

Exposes the model to diverse distortions, preventing intra-instance feature drift while preserving augmentation diversity.

Mean-Teacher Architecture

Anchors augmented representations toward a stable feature reference, reinforced by supervised contrastive loss.

Preference Data Construction

Constructs chosen-rejected explanation pairs to fine-tune the model to prioritize genuine manipulation evidence over superficial reasoning.

Balanaced Reward for concise explanation learning

Optimizes concise explanations via Group Relative Policy Optimization with rewards on semantic fidelity and conciseness.

Results

Experimental Results

We evaluate our framework on detection accuracy and explanation quality metrics, achieving state-of-the-art performance.

Detection Performance

Method	Accuracy
DINOv3 + MLP	0.8823
Ours	0.9824

Explanation Quality

Explain Model	SLEScore (norm/raw)	SimpleBert	SimpleOverall	ComplexBert	Entity F1	Claim F1
Baseline (Qwen3-VL-8B-Instruct)	0.1037 / -0.4812	0.5775	0.4354	0.5038	0.4402	0.3362
Ours	0.9939 / 3.9699	0.6930	0.7783	0.7241	0.5969	0.5227

Results

Challenge Results

We rank first among all participating teams on both the public leaderboard and the final hidden-test results of the ACM MM 2026 Explainable Deepfake Detection Challenge.

Challenge Top-5 Teams — Public Leaderboard

Top-5 teams on the public leaderboard of the challenge

Challenge Top-5 Teams — Final Hidden-Test Results

Pixel Sleuth, our team, ranks #1 on both the public leaderboard and the final hidden-test results.

Results

Qualitative Results

Deepfake Explanation Visualization

Visualizations for deepfake explanation. Baseline Qwen3-VL-8B-Instruct produces results with irrelevant information and fails to identify valid evidence. The SFT-only model captures partial evidence (marked in green) but still suffers from hallucinated irrelevant details (marked in red). Our model generates more complete and accurate explanation that identifies all manipulation traces (marked by green) without introducing spurious information.

Detection Robustness under Image Degradations

Qualitative comparison of detection robustness under various image degradations

Qualitative comparison of detection robustness under samples with various image degradations. The baseline model misclassifies degraded samples, while our model shows correct predictions across different degradation types.

Acknowledgements

We thank the following projects: vLLM, ms-swift, XPlainVerse.

Citation

BibTeX

@inproceedings{xu2026edd,
  title     = {Explainable Deepfake Detection with Feature-robust Augmentation and Evidence-grounded Explanation Optimization},
  author    = {Xu, Zhu and Yang, Yunfan and Peng, Yuxin and Liu, Yang},
  booktitle = {ACM International Conference on Multimedia},
  year      = {2026}
}