SuSy is a Spatial-Based Synthetic Image Detection and Recognition Model, designed and trained to detect synthetic images and attribute them to a generative model (i.e., two StableDiffusion models, two Midjourney versions and DALL·E 3). The model takes image patches of size 224x224 as input, and outputs the probability of the image being authentic or having been created by each of the aforementioned generative models.
image
The model is based on a CNN architecture and is trained using a supervised learning approach. It's design is based on previous work, originally intended for video superresolution detection, adapted here for the tasks of synthetic image detection and recognition. The architecture consists of two modules: a feature extractor and a multi-layer perceptron (MLP), as it's quite light weight. SuSy has a total of 12.7M parameters, with the feature extractor accounting for 12.5M parameters and the MLP accounting for the remaining 197K.
The CNN feature extractor consists of five stages following a ResNet-18 scheme. The output of each of the blocks is used as input for various bottleneck modules that are arranged in a staircase pattern. The bottleneck modules consist of three 2D convolutional layers. Each level of bottlenecks takes input at a later stage than the previous level, and each bottleneck module takes input from the current stage and, except the first bottleneck of each level, from the previous bottleneck module.
The outputs of each level of bottlenecks and stage 4 are passed to a 2D adaptative average pooling layer and then concatenated to form the feature map feeding the MLP. The MLP consists of three fully connected layers with 512, 256 and 256 units, respectively. Between each layer, a dropout layer (rate of 0.5) prevents overfitting. The output of the MLP has 6 units, corresponding to the number of classes in the dataset (5 synthetic models and 1 real image class).
The model can be used as a detector by either taking the class with the highest probability as the output or summing the probabilities of the synthetic classes and comparing them to the real class. The model can also be used as an recognition model by taking the class with the highest probability as the output.
This model can be used to detect synthetic images in a scalable manner, thanks to its small size. Since it operates on patches of 224x224, a moving window should be implemented in inference when applied on larger inputs (the most likely scenario, and the one it was trained under). This also enables the capacity for synthetic content localization within a high resolution input.
Any individual or organization seeking for support on the identification of synthetic content can use this model. However, it should not be used as the only source of evidence, particularly when applied to inputs produced by generative models not included in its training (see details in Training Data below).
Intended Uses
Intended uses include the following:
Detection of authentic and synthetic images
Attribution of synthetic images to their generative model (if included in the training data)
Localization of image patches likely to be synthetic or tampered.
Out-of-Scope Uses
Out-of-scope uses include the following:
Detection of manually edited images using traditional tools.
Detection of images automatically downscaled and/or upscaled. These are considered as non-synthetic samples in the model training phase.
Detection of inpainted images.
Detection of synthetic vs manually crafted illustrations. The model is trained mainly on photorealistic samples.
Attribution of synthetic images to their generative model if the model was not included in the training data. AThis model may not be used to train generative models or tools aimed at lthough some generalization capabilities are expected, reliability in this case cannot be estimated.
Forbidden Uses
This model may not be used to train generative models or tools aimed at purposefully deceiving the model or creating misleading content.
Bias, Risks, and Limitations
The model may be biased in the following ways:
The model may be biased towards the training data, which may not be representative of all authentic and synthetic images. Particularly for the class of real world images, which were obtained from a single source.
The model may be biased towards the generative models included in the training data, which may not be representative of all possible generative models. Particularly new ones, since all models included were released between 2022 and 2023.
The model may be biased towards certain type of images or contents. While it is trained using roughly 18K synthetic images, no assessment was made on which domains and profiles are included in those.
The model has the following technical limitations:
The performance of the model may be influenced by transformations and editions performed on the images. While the model was trained on some alterations (blur, brightness, compression and gamma) there are other alterations applicable to images that could reduce the model accuracy.
The performance of the model might vary depending on the type and source of images
The model will not be able to attribute synthetic images to their generative model if the model was not included in the training data.
The model is trained on patches with high gray-level contrast. For images composed entirely by low contrast regions, the model may not work as expected.
We use a random subset of the COCO dataset, containing 5,435 images, for the authentic images in our training dataset. The partitions are made respecting the original COCO splits, with 2,967 images in the training partition and 1,234 in the validation and test partitions.
For the diffusiondb dataset, we use a random subset of 5,435 images, with 2,967 in the training partition and 1,234 in the validation and test partitions. We use only the realistic images from the realisticSDXL dataset, with images in the realistic-2.2 split in our training data and the realistic-1 split for our test partition. The remaining datasets are used in their entirety, with 60% of the images in the training partition, 20% in the validation partition and 20% in the test partition.
To prepare the training data, we extract 240x240 patches from the images, minimizing the overlap between them. We then select the most informative patches by calculating the gray-level co-occurrence matrix (GLCM) for each patch. Given the GLCM, we calculate the contrast and select the five patches with the highest contrast. These patches are then passed to the model in their original RGB format and cropped to 224x224.
Recall: The proportion of correctly classified positive instances out of all actual positive instances in a dataset.
Results
Authentic Sources
Dataset
Year
Recall
Flickr30k
2014
90.53
Google Landmarks v2
2020
64.54
In-the-wild
2024
33.06
Synthetic Sources
Dataset
Model
Year
Recall
Synthbuster
Glide
2021
53.50
Synthbuster
Stable Diffusion 1.3
2022
87.00
Synthbuster
Stable Diffusion 1.4
2022
87.10
Synthbuster
Stable Diffusion 2
2022
68.40
Synthbuster
DALL-E 2
2022
20.70
Synthbuster
MidJourney V5
2023
73.10
Synthbuster
Stable Diffusion XL
2023
79.50
Synthbuster
Firefly
2023
40.90
Synthbuster
DALL-E 3
2023
88.60
Authors
Stable Diffusion 3 Medium
2024
93.23
Authors
Flux.1-dev
2024
96.46
In-the-wild
Mixed/Unknown
2024
89.90
Summary
The results for authentic image datasets reveal varying detection performance across different sources. Recall rates range from 33.06% for the In-the-wild dataset to 90.53% for the Flickr30k dataset. The Google Landmarks v2 dataset shows an intermediate recall rate of 64.54%. These results indicate a significant disparity in the detectability of authentic images across different datasets, with the In-the-wild dataset presenting the most challenging case for SuSy.
The results for synthetic image datasets show varying detection performance across different image generation models. Recall rates range from 20.70% for DALL-E 2 (2022) to 96.46% for Flux.1-dev (2024). Stable Diffusion models generally exhibited high detectability, with versions 1.3 and 1.4 (2022) showing recall rates above 87%. More recent models tested by the authors, such as Stable Diffusion 3 Medium (2024) and Flux.1-dev (2024), demonstrate even higher detectability with recall rates above 93%. The in-the-wild mixed/unknown synthetic dataset from 2024 showed a high recall of 89.90%, indicating effective detection across various unknown generation methods. These results suggest an overall trend of improving detection capabilities for synthetic images, with newer generation models generally being more easily detectable.
It must be noted that these metrics were computed using the center-patch of images, instead of using the patch voting mechanisms described previously. This strategy allows a more fair comparison with other state-of-the-art methods although it hinders the performance of SuSy.
Environmental Impact
Hardware Type: H100
Hours used: 16
Hardware Provider: Barcelona Supercomputing Center (BSC)
Compute Region: Spain
Carbon Emitted: 0.63kg
Citation
BibTeX:
bibtex
1@misc{bernabeu2024susy,
2 title={Present and Future Generalization of Synthetic Image Detectors},
3 author={Pablo Bernabeu-Perez and Enrique Lopez-Cuena and Dario Garcia-Gasulla},
4 year={2024},
5 eprint={2409.14128},
6 archivePrefix={arXiv},
7 primaryClass={cs.CV},
8 url={https://arxiv.org/abs/2409.14128},
9}
bibtex
1@thesis{bernabeu2024aidetection,
2 title={Detecting and Attributing AI-Generated Images with Machine Learning},
3 author={Bernabeu Perez, Pablo},
4 school={UPC, Facultat d'Informàtica de Barcelona, Departament de Ciències de la Computació},
5 year={2024},
6 month={06}
7}
