James Rainey Artificial Intelligence has transformed how we create and edit images, for better and for worse. While tools like DALL-E and Stable Diffusion enable creative expression, they also blur the line between real and fake.
Goal
This research takes on the challenge of AI manipulated images by developing:
-
A large-scale dataset of AI-manipulated images
-
A deep learning detection model that identifies and localizes AI-based edits
Our Contributions
- The NCL-IMD.v1 Dataset
We built a dataset of 14,000+ AI-manipulated images, each paired with ground truth masks that show exactly where alterations occurred.
Created using Blended Latent Diffusion, a state-of-the-art text-guided inpainting model.
Includes a wide range of image contexts, mask sizes, and manipulation types.
Publicly available: MediaTrust Dataset
- A New Detection Model
We proposed a CNN-based Autoencoder architecture designed specifically to detect manipulations made by diffusion models — where existing models (like MM-Fusion) fail.
Outperforms MM-Fusion on AI-generated images.
F1 score improved from 0.27 → 0.50, with high accuracy (≈96%).
How We Built the Dataset
The data generation process was fully automated:
-
Mask Generator – creates random, human-like selection areas on images.
-
Prompt Generator – uses CLIP + GPT4All (Mistral Instruct) to generate text prompts describing modifications.
-
Image Manipulation – applies Blended Latent Diffusion to generate localized edits.
-
Balancing & Splitting – ensures diverse, balanced data across train/validation/test sets.
Total output: 13,614 high-resolution manipulated images Each image has an associated binary mask indicating the edited regions.
Model Development
Existing detectors (like TruFor and MM-Fusion) perform well on traditional photo edits but struggle with AI-generated content — especially diffusion-based manipulations.
We designed a VGG19-based Autoencoder optimized for anomaly detection in image space.
Removed weaker modalities (e.g., Noiseprint++)
Trained on the NCL-IMD.v1 dataset
Used Mean Squared Error (MSE) loss for stable convergence
The result was a faster, more reliable, and explainable model for identifying AI edits.
Results
Model F1 Accuracy Precision Recall MM-Fusion (default) 0.28 0.85 0.33 0.71 MM-Fusion (retrained) 0.23 0.65 0.16 0.56 CNN-Autoencoder (ours) 0.50 0.96 0.53 0.55
Our Autoencoder model more than doubled the detection performance on diffusion-generated images compared to existing approaches.
Why It Matters
TODO
Key Takeaways
Diffusion-based image edits are fundamentally harder to detect than GAN-based ones.
Prompt quality affects realism — weak text inputs can cause unrealistic or stylistic artifacts.
A data-centric approach (building large, balanced datasets) is crucial for progress in forensic AI research.
Looking Ahead
Improve realism of generated manipulations using next-gen diffusion models.
Refine prompt generation for context-aware edits.
Combine visual and semantic features for hybrid detection.
Develop open benchmarks for evaluating AI manipulation detection globally.