🤖 AI

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

This paper proposes a novel framework for explainable deepfake detection that combines feature-robust augmentation with supervised contrastive learning to enhance resilience against image degradation, and employs evidence-grounded preference optimization to generate factually accurate explanations, achieving first place in the ACM Multimedia 2026 Explainable Deepfake Detection Challenge.

Original authors: Zhu Xu, Jiaqi Tang, Pokai Chen, Yuxin Peng, Yang Liu

Published 2026-08-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhu Xu, Jiaqi Tang, Pokai Chen, Yuxin Peng, Yang Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital age, the line between what is real and what is manufactured has become increasingly blurred. Generative artificial intelligence has advanced to the point where it can create images that are nearly indistinguishable from photographs taken by a camera. These fabricated images, often called deepfakes, pose a serious threat to the integrity of news, legal evidence, and personal trust. For years, researchers have focused on building computer programs that can simply say "yes" or "no" when shown an image, deciding whether it is authentic or fake. However, in the real world, a simple verdict is rarely enough. A forensic expert or a fact-checker needs to know why an image is suspicious. They need to see the specific visual clues—the unnatural shadows, the distorted textures, or the impossible geometry—that reveal the deception. This need has given rise to a new field of study: explainable detection, where the goal is not just to find the fake, but to explain the evidence clearly to a human observer.

Despite progress in this area, existing systems have struggled with two major problems that limit their usefulness. First, these systems are often fragile; if an image is slightly blurry, compressed, or of low quality, the detector's accuracy can collapse. Second, when these systems do try to explain their reasoning, they often fail to tell the truth. They might miss the actual evidence of manipulation or, worse, invent details that do not exist, a behavior known as hallucination. A team of researchers from Peking University has addressed these issues with a new framework that won first place in the 2026 ACM Multimedia Explainable Deepfake Detection Challenge. Their approach does not just improve the ability to spot fakes; it ensures the system remains reliable even when image quality is poor and forces the explanation to stick strictly to the factual evidence found in the picture.

The researchers began by tackling the problem of image quality. In the real world, images are rarely perfect; they are often resized, compressed, or degraded by noise. The team found that standard methods of training computers to recognize fakes often backfire when exposed to too many different types of image distortions. Instead of learning to recognize the fake, the computer gets confused by the variety of changes, a phenomenon the authors describe as feature drift. To fix this, they developed a training method that exposes the model to a wide range of degraded images while simultaneously teaching it to hold onto a stable understanding of what a real or fake image looks like. They used a technique where a "teacher" model, which is a smoothed, stable version of the learning model, guides the student. This ensures that no matter how much the image is distorted, the computer's internal representation of the image stays consistent. This allows the system to detect deepfakes accurately even when the image quality is poor, a capability that previous models lacked.

Once the system can reliably detect a fake, the next challenge is to explain why. The researchers observed that standard explanation models often produce answers that sound plausible but are factually wrong. They might focus on irrelevant details, like the color of a shirt, while ignoring the actual manipulation, such as a warped hand. To solve this, the team created a new training process that teaches the model to value truth over style. They constructed a dataset where the computer was shown pairs of explanations: one that was complete and accurate, and another that was flawed. The flawed examples were specifically designed to mimic common mistakes, such as leaving out crucial evidence or adding made-up details. By training the model to prefer the accurate explanations and reject the flawed ones, the system learned to prioritize genuine manipulation traces. This process, known as preference optimization, effectively filtered out the hallucinations and forced the model to ground its reasoning in the actual visual evidence present in the image.

The result is a system that is both robust and honest. In tests using a massive dataset of one million images, the new framework outperformed all other competitors, including baseline models from major technology companies. It achieved the highest scores in both detecting deepfakes and generating explanations that were semantically faithful to the evidence. The researchers also developed a second, streamlined version of the model for situations where speed and brevity are more important than detailed analysis. This version was trained to provide concise summaries of the most critical clues, balancing the need for accuracy with the need for quick readability. The success of this work demonstrates that it is possible to build artificial intelligence that not only sees the truth but can also explain it clearly, without being misled by poor image quality or its own tendency to invent facts. This advancement offers a more reliable tool for forensic analysts and the public alike, helping to restore trust in digital media by making the detection process transparent and evidence-based.

Technical Summary: Explainable Deepfake Detection with Feature-robust Augmentation and Evidence-grounded Explanation Optimization

Problem Statement
The paper addresses the limitations of current explainable deepfake detection systems, which extend beyond binary classification to require interpretable justifications. The authors identify two critical deficiencies in existing approaches that hinder real-world applicability:

Vulnerability to Image Quality Degradation: Detection accuracy significantly declines on low-quality samples. While standard data augmentation is a common remedy, the authors observe that excessively diverse augmentation strategies can induce "feature drift," where the model struggles to learn consistent representations, causing classification boundaries to blur and performance to drop.
Factually Flawed Explanations: Current explanation models often fail to provide accurate rationales. Empirical analysis reveals two prevalent failure modes: (i) Evidence Omission, where crucial manipulation traces are missing (observed in 68.3% of baseline samples), and (ii) Irrelevant Information Intrusion, where the model hallucinates spurious details (observed in 57.7% of samples). Existing methods often focus on the procedural format of reasoning (e.g., length or step count) rather than the factual accuracy and completeness of the evidence.

Methodology
The authors propose a unified framework comprising two synergistic components: a robust detection branch and an evidence-grounded explanation branch.

Feature-robust Augmentation for Detection:
- Architecture: The system utilizes DINOv3-7B as a visual encoder, aggregating multi-granularity tokens (CLS, REG, and AVG patch tokens) to capture both global semantics and fine-grained details.
- Degradation-Aware Augmentation Pipeline: To simulate real-world conditions, the training pipeline applies diverse distortions (smoothing, resizing, Gaussian/Non-Gaussian noise, JPEG compression, color enhancement, and random distractors) with varying intensities.
- Supervised Contrastive Learning: A supervised contrastive loss is employed to pull features of the same authenticity class (real or fake) into compact clusters regardless of degradation, enhancing discriminability.
- Mean-Teacher Stabilization: To prevent intra-instance feature drift caused by diverse augmentations, a mean-teacher architecture is introduced. The teacher network, updated via an Exponential Moving Average (EMA) of the student's weights, provides a stable feature anchor. A consistency loss aligns the student's features from degraded views with the teacher's features from the clean view, ensuring stability without sacrificing augmentation diversity.
Evidence-grounded Explanation Optimization:
- Preference Dataset Construction: The authors construct a preference dataset ( $D_{pref}$ ) consisting of "chosen-rejected" explanation pairs. For each ground-truth explanation, rejected samples are synthetically generated by an advanced MLLM (Qwen3.5-27B) via three strategies: (i) removing key manipulation traces (Evidence-Omitted), (ii) injecting irrelevant but visually present details (Irrelevant-induced), and (iii) a combination of both.
- Direct Preference Optimization (DPO): The model is fine-tuned using DPO to prioritize chosen (complete and accurate) explanations over rejected ones. This explicitly teaches the model to avoid evidence omission and hallucination.
- Concise Explanation Model: For scenarios requiring brevity, a separate model is optimized via Group Relative Policy Optimization (GRPO). This model uses a reward function balancing semantic fidelity (measured by BERTScore) and conciseness (measured by SLEscore) to produce compact, critical cues.

Key Contributions

Empirical Validation of Vulnerabilities: The paper provides empirical evidence demonstrating that detection accuracy is highly sensitive to image quality degradation and that standard augmentation can degrade performance due to feature drift. It also quantifies the prevalence of evidence omission and hallucination in existing explanation models.
Novel Framework: The proposed framework jointly enhances detection robustness through degradation-aware augmentation paired with mean-teacher stabilization and improves explanation faithfulness through evidence-grounded preference optimization.
State-of-the-Art Performance: The approach achieves first place in the ACM Multimedia 2026 Explainable Deepfake Detection Challenge, demonstrating superior performance in both detection robustness and explanation interpretability.

Results
Evaluated on the XPlainVerse test set (200k images), the proposed method ("Pixel Sleuth") outperformed 138 participating teams:

Overall Score: Achieved the highest overall score (0.841617 on the public leaderboard; 0.7612 on the hidden test set).
Detection: Achieved a detection F1 score of 0.9424 on the hidden test set, surpassing the baseline Qwen3-VL-8B (0.6342) and other top teams.
Explanation Quality: Secured the highest scores across all explanation metrics, including Complex BERTScore (0.7063), Simple BERTScore (0.6819), Entity F1 (0.4205), and Claim F1 (0.5800).
Qualitative Analysis: Visualizations confirm that the model maintains accurate predictions under various degradations where baselines fail, and generates explanations that accurately enumerate manipulation traces without fabricating non-existent artifacts.

Significance
The paper claims that its framework offers a reliable solution for explainable deepfake detection by addressing the dual challenges of robustness against quality variations and the factual accuracy of explanations. By shifting the focus from procedural reasoning formats to evidence-grounded optimization, the work facilitates more effective forensic scrutiny and fosters greater user confidence in automated detection tools. The inclusion of both detailed forensic analysis and concise fact-checking models allows the system to adapt flexibly to diverse usage scenarios.

Technical Summary: Explainable Deepfake Detection with Feature-robust Augmentation and Evidence-grounded Explanation Optimization

More like this