Introduction
AI-generated synthetic images have rapidly transformed the digital landscape. With text-to-image (T2I) models evolving at an incredible pace, distinguishing fake from real images has become increasingly difficult. While these advancements open creative possibilities, they also enable identity theft, misinformation, and fraud.
To counter these threats, researchers introduced DeepGuard—a powerful three-level detection framework. This system not only detects AI-generated images but also attributes them to the models that produced them. As a result, DeepGuard offers a promising solution to restore trust in digital media.
The Methodology Behind Detecting AI-Generated Synthetic Images
DeepGuard operates on a three-tiered framework that integrates deep learning, ensemble learning, and model attribution. Each level contributes uniquely to detecting and identifying synthetic images with high accuracy.
Level 0: Binary Classification of Synthetic Images
In the first stage, DeepGuard uses state-of-the-art deep learning models to distinguish between real and synthetic images. Models like DenseNet201, DenseNet121, ResNet34, and EfficientNet-B3 undergo rigorous training on the DeepGuardDB dataset. The system evaluates image artifacts, texture inconsistencies, and resolution differences to flag manipulated visuals.
To prevent overfitting, the team applied early stopping during training and used 5-fold cross-validation, which improved generalization. By ensuring high initial accuracy, Level 0 sets a strong foundation for subsequent classification layers.
Level 1: Ensemble Learning for Meta-Classification
Rather than relying on a single model, DeepGuard enhances its decision-making using ensemble learning. At this level, the outputs from Level 0’s top models—DenseNet201, DenseNet121, and ResNet34—are aggregated using ensemble methods such as Logistic Regression, AdaBoost, and Random Forest.
These ensemble techniques combine probabilities to generate final predictions. By leveraging the strengths of individual models, DeepGuard reduces both false positives and negatives. As a result, this layer improves detection consistency across various image categories.
Level 2: Multi-Class Model Attribution
In the final stage, DeepGuard classifies detected synthetic images according to their source generation models, such as Stable Diffusion 3, DALL-E 3, and Imagen. This step uses multi-class classifiers trained on labeled subsets of AI-generated content from the DeepGuardDB dataset.
Models like ResNet50 and ResNet34 deliver impressive accuracy. They assign fake images to their correct generators by identifying subtle stylistic and resolution features unique to each platform. This level ensures that AI misuse can be traced to specific tools, enhancing accountability.
How DeepGuard Works: The Process Behind Detecting Synthetic Images
Synthetic Images: A Balanced and Diverse Dataset
DeepGuard’s backbone is its robust dataset, DeepGuardDB, comprising 13,000 balanced samples—50% real and 50% AI-generated. Researchers sourced real images from MS-COCO and Flickr30k, ensuring a wide range of scenes, from rural landscapes to human interactions.
For synthetic images, the team used top-tier T2I tools like Stable Diffusion 3, Imagen, DALL-E 3, and GLIDE. Importantly, they employed the same prompts across both real and synthetic image creation. This method ensures contextual similarity, which strengthens model training and evaluation.
Table 2. Overview of Fake Images in DeepGuardDB
| Model | Dataset Source | Generated Images | Image Size |
|---|---|---|---|
| Stable Diffusion | MS-COCO | 1337 | 512×512 |
| Stable Diffusion | Flickr30k | 1338 | 512×512 |
| Imagen | MS-COCO | 588 | 512×512 |
| Imagen | Flickr30k | 587 | 512×512 |
| DALL-E3 | MS-COCO | 2150 | 512×512 |
| DALL-E3 | Flickr30k | 2150 | 512×512 |
| GLIDE | MS-COCO | 250 | 512×512 |
| GLIDE | Flickr30k | 250 | 512×512 |
Results: Accuracy and Effectiveness in Identifying Synthetic Images
Synthetic Images: Binary Classifier Performance (Level 0)
DeepGuard initially tested eight binary classifiers. Among them, DenseNet201 performed best, achieving 99.83% accuracy, followed closely by ResNet34 and DenseNet121. These models demonstrated excellent precision, recall, and F1-scores, highlighting their ability to detect fine-grained visual differences.
Table 3. Binary Classifier Performance
| Model | Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|
| DenseNet201 | 99.83% | 99.80% | 99.86% | 99.83% |
| ResNet34 | 99.53% | 99.73% | 99.33% | 99.54% |
| DenseNet121 | 99.00% | 98.33% | 99.66% | 98.99% |
| EfficientNet-B3 | 98.30% | 98.60% | 98.01% | 98.30% |
Meta-Classifier Results (Level 1)
When combining the top three Level 0 models, ensemble methods significantly boosted accuracy. All three meta-classifiers—Logistic Regression, AdaBoost, and Random Forest—exceeded 99.87% accuracy. This result proves that blending predictions refines output quality.
Table 4. Meta-Classifier Results
| Meta Model | Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|
| Logistic Regression | 99.87% ± 0.19 | 99.93% ± 0.13 | 99.80% ± 0.40 | 99.87% ± 0.20 |
| AdaBoost | 99.90% ± 0.13 | 99.93% ± 0.13 | 99.87% ± 0.27 | 99.90% ± 0.13 |
| Random Forest | 99.90% ± 0.13 | 99.93% ± 0.13 | 99.87% ± 0.27 | 99.90% ± 0.13 |
Model Attribution (Level 2)
The final layer attributed images to their generating tools. Both ResNet34 and ResNet50 achieved over 92% accuracy, with ResNet50 showing slightly better general performance. Although precision was slightly lower for DALL-E and Imagen, results remained robust.
Misclassifications in these two categories stem from ethical constraints within DALL-E and data limitations in Imagen, which often produce less-detailed human features.
Why Identifying Synthetic Images Matters
As generative AI grows, synthetic images become harder to detect. These visuals can manipulate elections, falsify identities, or fabricate legal evidence. Consequently, detection systems must evolve to safeguard truth in digital media.
DeepGuard serves this mission by:
- Preventing misinformation in journalism and social media
- Supporting legal investigations with verifiable evidence
- Securing digital platforms against fake content
- Enhancing transparency by attributing fake images to source models
Future Research and Improvements in AI-Generated Image Detection
Moving forward, the research team aims to:
- Expand DeepGuardDB with challenging scenarios like low-resolution and occluded images
- Explore real-time detection APIs for integration into forensic and surveillance systems
- Combine multi-modal detection, merging image data with metadata and text
- Extend model capabilities to detect deepfake videos and 3D content
Through these developments, DeepGuard will remain adaptable as AI-generated content continues to evolve.
Conclusion
DeepGuard represents a major step forward in synthetic image detection. Its robust three-level architecture accurately flags fake content and traces it to its origin, maintaining both transparency and accountability.
By blending technical sophistication with practical applicability, DeepGuard empowers industries to combat visual misinformation and protect digital integrity. As AI continues to reshape how we perceive reality, tools like DeepGuard will play a critical role in preserving the truth.
Reference: Namani, Y.; Reghioua, I.; Bendiab, G.; Labiod, M.A.; Shiaeles, S. DeepGuard: Identification and Attribution of AI-Generated Synthetic Images. Electronics 2025, 14, 665. https://doi.org/10.3390/electronics14040665
CC BY 4.0 License: This paper is distributed under the Creative Commons Attribution 4.0 International License (CC BY 4.0), which allows unrestricted use, distribution, and reproduction in any medium, provided that the original work is properly credited. You can read the full terms here: https://creativecommons.org/licenses/by/4.0/.