|
--- |
|
license: unknown |
|
metrics: |
|
- accuracy |
|
tags: |
|
- art |
|
datasets: |
|
- DataScienceProject/Art_Images_Ai_And_Real_ |
|
library_name: keras |
|
language: |
|
- en |
|
pipeline_tag: image-classification |
|
--- |
|
|
|
# Model Card for Model ID |
|
|
|
This model is designed for classifying images as either 'real' or 'fake-AI generated' using a Convolutional Neural Network (CNN) combined with Error Level Analysis (ELA). |
|
|
|
Our goal is to accurately classify the source of the image with at least 85% accuracy and achieve at least 80% in the recall test. |
|
|
|
|
|
|
|
|
|
### Model Description |
|
|
|
CNN is a type of deep learning model specifically designed to process and analyze visual data by applying convolutional layers that automatically detect patterns and features in images. |
|
|
|
Error Level Analysis (ELA) detects changes in digital images by comparing the differences between the original and a compressed version of the image. |
|
It highlights areas where alterations may have occurred, making it useful for identifying image tampering. |
|
|
|
After running ELA we feed the CNN with the result of comparing original and a compressed version of the same image and then we get the output. |
|
|
|
### Direct Use |
|
|
|
This model can be used to classify images as 'real' or 'fake- Ai generated' based on the presence of anomalies and features characteristic of each category. |
|
|
|
|
|
### Out-of-Scope Use |
|
|
|
The model may not perform well on images outside the scope of art or where the visual characteristics are drastically different from those in the training dataset. |
|
|
|
may not detect ai-images from newer diffusion models that has another method of creating ai images. |
|
|
|
## Bias, Risks, and Limitations |
|
|
|
ELA can be very good in detecting tamparing such as photoshop, |
|
there is many diffusion engines that not tamper the image making the ELA weak against those ai-images. |
|
|
|
### Recommendations |
|
|
|
Test this model with different hyperparameters / more CNN layers |
|
|
|
## How to Get Started with the Model |
|
|
|
Prepare Data: Organize your images into appropriate folders and run the code |
|
|
|
|
|
## Training Details |
|
|
|
|
|
-Dataset: DataScienceProject/Art_Images_Ai_And_Real_ |
|
|
|
-Preprocessing: Images are resized, image quallity changed , ELA version of this image created. |
|
|
|
### Training Procedure |
|
|
|
Train the CNN with the Preprocessed images , use valitadion set. |
|
|
|
|
|
#### Training Hyperparameters |
|
|
|
|
|
-optimizer = RMSprop(lr=0.0005, rho=0.9, epsilon=1e-08, decay=0.0) |
|
|
|
-epochs = 22 |
|
|
|
-batch_size = 100 |
|
|
|
-loss = "categorical_crossentropy" |
|
|
|
-metrics=["accuracy"] |
|
|
|
-early_stopping = EarlyStopping(monitor='val_acc',min_delta=0,patience=2,verbose=0, mode='auto') |
|
|
|
|
|
|
|
## CNN architecture |
|
 |
|
|
|
|
|
## Evaluation |
|
|
|
The model takes 7-10 minutes to run , based on our dataset , equipped with the following pc hardware: cpu :i9 13900 ,ram: 32gb , gpu: rtx 3080 |
|
your mileage may vary. |
|
|
|
|
|
### Testing Data, Factors & Metrics |
|
|
|
-precision |
|
-recall |
|
-f1 |
|
-confusion_matrix |
|
-accuracy |
|
|
|
### Results |
|
-test accuracy = 0.853 |
|
|
|
-precision = 0.866 |
|
|
|
-recall = 0.8357 |
|
|
|
-f1 = 0.8509 |
|
|
|
|
|
|
|
|
|
 |
|
|
|
#### Summary |
|
|
|
The model preforms well and meet our initial goal ,thus this model can handle the task of image classification real image vs Ai generated image. |
