DFBench: Benchmarking Deepfake Image Detection Capability of Large Multimodal Models

Jiarui Wang, Huiyu Duan, Juntong Wang, Ziheng Jia, Woo Yi Yang, Xiaorong Zhu, Yu Zhao, Jiaying Qian, Yuke Xing, Guangtao Zhai, Xiongkuo Min

Abstract

As generative AI transitions from the "uncanny valley" to producing imagery indistinguishable from reality, existing deepfake detection tools face significant challenges in generalization and robustness. This paper introduces DFBench, a comprehensive benchmark comprising 540,000 images designed to evaluate Large Multimodal Models (LMMs) in deepfake detection. We further propose MoA-DF, a "Mixture of Agents" framework that aggregates the predictive power of multiple LMMs. Our results demonstrate that LMMs, particularly when fine-tuned via Low-Rank Adaptation (LoRA), significantly outperform specialized detectors in handling real-world distortions and identifying sophisticated AI-generated content, achieving a state-of-the-art accuracy of 98.41% on real image subsets.

1. Introduction

The rapid evolution of generative AI has outpaced the development of reliable detection mechanisms. Current research in deepfake detection is hindered by three primary limitations: outdated benchmarks that focus almost exclusively on human faces or older GAN models; a lack of real-world complexity, such as motion blur and compression artifacts; and the underutilization of Large Multimodal Models (LMMs) which possess inherent visual reasoning capabilities.

DFBench addresses these gaps by providing a massive, diverse dataset that includes not only fully synthetic images but also partially edited content and distorted real-world photography. By benchmarking LMMs against traditional specialized detectors, this work highlights the potential of general-purpose AI as a robust defense against increasingly deceptive synthetic media.

2. The DFBench Dataset

DFBench is a large-scale dataset consisting of 540,000 images, categorized to reflect the complexities of modern digital environments:

  • Real Images (45,000): Sourced from eight public repositories (e.g., Flickr8k, TID2013), these include natural distortions like noise and blur to simulate real-world conditions.
  • AI-Edited Images (15,000): Derived from EPAIQA-15K, these images feature localized modifications such as object enhancement, style changes, and semantic operations, which are often harder to detect than full generation.
  • AI-Generated Images (480,000): This subset spans 12 state-of-the-art models, including diffusion-based models (Flux, SD3.5-Large, PixArt-sigma) and autoregressive models (Janus, Infinity).

3. Methodology: MoA-DF

To maximize detection efficacy, the researchers developed MoA-DF (Mixture of Agents for Deepfake Detection). This framework utilizes an ensemble of three high-performing LMMs: Qwen2.5 (7B), InternVL2.5 (8B), and InternVL3 (9B).

Each model acts as an independent agent, providing a probability score regarding the authenticity of an image. These scores are aggregated to produce a final decision, mitigating the biases or errors of any single model. To optimize performance, the agents were fine-tuned using LoRA (Low-Rank Adaptation) on a subset of the data. Remarkably, fine-tuning for just one epoch allowed these LMMs to surpass specialized models that had undergone significantly longer training durations.

4. Key Insights and Results

4.1 Resilience to the "Distortion Trap"

Traditional specialized detectors often suffer from high false-positive rates when encountering real images with low quality. These models frequently mistake natural blur or compression for AI artifacts. In contrast, LMMs demonstrate a superior understanding of image context, correctly identifying distorted real images as "real" while maintaining high sensitivity to actual synthetic patterns.

4.2 The Difficulty of AI Editing

The benchmark reveals that detecting partial AI edits is significantly more challenging than detecting fully synthetic images. While most models perform well on full generations, their accuracy drops when faced with subtle semantic changes or object operations within a real photograph.

4.3 Performance Metrics

MoA-DF achieved a dominant 98.41% accuracy on real image subsets. In comparative testing, while standard detectors like CnnSpott achieved high accuracy on pristine real images, their performance plummeted when exposed to AI-edited content. MoA-DF remained stable across all categories. Notably, SD3.5-Large was identified as the "Evasion King," producing the most deceptive images that consistently resulted in the lowest detection rates across all tested models.

5. Discussion and Limitations

While LMMs represent a significant leap forward in detection capability, they introduce new challenges. The computational cost of running an ensemble of three large models is substantially higher than that of lightweight CNNs, posing a barrier to real-time, high-volume applications. Furthermore, the "arms race" between generators and detectors continues; as models like Flux and SD3.5 evolve, detection benchmarks must be continuously updated. Future work should also investigate prompt sensitivity, as the phrasing of the detection query can influence LMM output.

6. Conclusion

DFBench and MoA-DF establish a new standard for deepfake detection research. By leveraging the broad reasoning capabilities of Large Multimodal Models, we can move beyond the limitations of specialized, narrow detectors and build systems capable of navigating the complexities of modern, AI-augmented visual media.

Summary updated

BibTeX
@inproceedings{wang2025dfbench,
  author = {Jiarui Wang and Huiyu Duan and Juntong Wang and Ziheng Jia and Woo Yi Yang and Xiaorong Zhu and Yu Zhao and Jiaying Qian and Yuke Xing and Guangtao Zhai and Xiongkuo Min},
  title = {{DFBench: Benchmarking Deepfake Image Detection Capability of Large Multimodal Models}},
  booktitle = {ACM Multimedia},
  year = {2025},
  doi = {10.1145/3746027.3758204},
}
Figure