FIND: A Simple yet Effective Baseline for Diffusion-Generated Image Detection

1. Problem Statement

The rapid advancement of diffusion models (e.g., Stable Diffusion, Midjourney) has enabled the generation of highly realistic images, posing significant challenges for digital media trust, misinformation detection, and copyright protection.

• Current Limitations: Existing detection methods (e.g., DIRE, LaRE2) rely on reconstruction errors. They use an additional diffusion model to reconstruct the input image and measure the difference between the original and the reconstructed version. The underlying assumption is that synthetic images are easier to reconstruct than real images.
• Drawbacks of Current Approaches:
  • High Computational Cost: Reconstruction requires running complex diffusion processes (often multiple steps), making inference slow.
  • Poor Generalizability: Performance is heavily dependent on the specific diffusion model used for reconstruction. If the detector is trained on one model's reconstruction logic, it often fails to generalize to images generated by unseen models.

2. Core Hypothesis & Methodology

The authors propose FIND (Forgery Identification via Noise Disturbance), a method that eliminates the need for explicit reconstruction and auxiliary diffusion models.

Theoretical Foundation

The paper hypothesizes that the discriminative power of reconstruction error stems from the difficulty of fitting image distributions with Gaussian distributions:

• Real Images: Composed of a complex superposition of a vast (potentially infinite) number of Gaussian components. They are difficult to approximate with a finite Gaussian Mixture Model (GMM), leading to high reconstruction errors.
• Synthetic Images: Generated via finite diffusion iterations, they represent a superposition of a finite number of Gaussian distributions. They are easier to fit with a GMM, resulting in lower reconstruction errors.

The FIND Approach

Instead of calculating reconstruction error explicitly, FIND implicitly teaches a binary classifier to distinguish these distributional differences by manipulating the training data:

1. Noise Augmentation: During training, random Gaussian noise is added to real images.
2. Labeling: These "noisy real images" are labeled as synthetic (fake).
3. Rationale:
   • Adding noise increases the variance of the real image distribution, smoothing it out and making it easier to fit with a Gaussian distribution (mimicking the statistical properties of synthetic images).
   • Crucially, the noisy images retain visual similarity to the original real images. This forces the classifier to ignore visual artifacts (which are similar in both noisy-real and real) and focus solely on the statistical distribution differences that distinguish true synthetic images from real ones.
4. Training: A standard binary classifier (e.g., ViT or ResNet) is trained on three types of data:
   • Original Real Images (Label: Real)
   • Noisy Real Images (Label: Synthetic)
   • Synthetic Images (Label: Synthetic)
5. Inference: During testing, images are fed directly into the classifier. No noise addition or reconstruction is performed.

3. Key Contributions

• Theoretical Insight: The paper provides a theoretical proof that reconstruction error is essentially a measure of how well an image distribution fits a Gaussian mixture model. It demonstrates that adding noise to real images makes them statistically similar to synthetic images in terms of "Gaussian fit-ability."
• Efficiency: FIND removes the reconstruction pipeline entirely. It requires only a simple forward pass through a classifier, eliminating the need for heavy diffusion models during inference.
• Generalizability: By avoiding model-specific reconstruction logic, FIND achieves superior cross-model generalization.
• Simplicity: The method requires only a standard binary classifier and a simple noise-augmentation strategy during training, serving as a strong, reproducible baseline.

4. Experimental Results

The method was evaluated on the GenImage benchmark, a large-scale dataset containing over 2.6 million images from 8 different generators (including Stable Diffusion v1.4/v1.5, Midjourney, ADM, GLIDE, etc.).

• Accuracy:
  • Using a ResNet50 backbone, FIND achieved 80.13% average accuracy, outperforming the previous SOTA (LaRE2 at 79.07%).
  • Using a ViT-B/16 backbone, FIND achieved 88.35% average accuracy, an 11.7% improvement over existing methods.
  • FIND showed consistent high performance across all 8 generators, whereas previous methods (DIRE, LaRE2) struggled significantly when the test generator differed from the reconstruction model.
• Efficiency (Speed):
  • FIND is 126× faster than existing methods.
  • Inference Time: FIND takes 0.4 ms per image (2500 FPS).
  • Comparison: DIRE takes ~151.9 ms (6.6 FPS), and LaRE2 takes ~50.4 ms (19.8 FPS).
• Ablation Studies:
  • Adding noise to real images (the core FIND strategy) was found to be the most critical factor for performance.
  • Adding noise to synthetic images or using pure Gaussian noise images as data was less effective or detrimental.
  • An optimal noise magnitude ($\epsilon$) of 50 was identified, balancing statistical smoothing with visual information retention.

5. Significance

• Paradigm Shift: FIND shifts the detection paradigm from "reconstruction-based" to "distribution-based," proving that complex generative models are not required to detect generative models.
• Practical Deployment: The extreme speed (2500 FPS) and lack of dependency on specific diffusion models make FIND highly suitable for real-time applications, such as content moderation on social media platforms or real-time video streaming.
• New Baseline: The paper establishes a new, simple, and highly effective baseline for the field of synthetic image detection, suggesting that future research should focus on distributional properties rather than expensive reconstruction pipelines.