DeepFakeCounter

Deep Fake Counter or "DFC" aims to bring a classification method to determining if an image was generated by AI or not.

Dataset

• CIFAKE: Real and AI-Generated Synthetic Images

How to Run

Create the conda environment

conda create -y -n deepfakecounter python=3.10

Activate the environment

conda activate deepfakecounter

Pip install the requirements.txt

pip install -r requirements.txt

Change directory into the src folder then run main.py

cd src/

Go into the config.py file. I'm assuming you're on a gpu. If not, change ACCELERATOR to 'cpu' then run the following

python main.py

Architectures

The winning architecture of the 4 I implemented, is as follows...

• Tail
  • 4 convolutions where after each convolution, we do the following
    • BatchNorm
    • MaxPool
    • RELU Activation
• Flatten
• Head
  • 2 Linear Layers
    • RELU Activation
  • Output Linear Layer
    • Sigmoid

You can find all the implementations in the src/models/cnn.py file. The model not commented is the 4_conv_batch_3_linear architecture.

Results

Architectures / Eval	Training	Test
2_conv_2_linear_paper	0.944	0.927
3_conv_3_linear	0.987	0.951
4_conv_batch_3_linear	0.984	0.952
5_conv_batch_3_linear	0.979	0.945

Result Analysis

I was perfectly able to replicate the CIFAKE paper's test accuracy results. While their proposal was good, I thought to use more convolutions and batch normalization. The reason is simply because when it comes to detecting fake pictures from the real ones, I thought that we needed to pay closer attention to finer details. When you're trying to figure out if this picture of a human has been deep faked, you can tell easily by the hair. Sometimes in fake images, the hair may be dull in texture - this is a dead giveaway. The same can be said about hands along with the background of the subject. Identifying these small pattern differences is what I believe is the "end all be all" for detecting fake images.

Digging for too small of detail may also hinder the performance hence why greater than 4 convolutions started to hinder performance for the model. Finding that sweet spot of 4 convolutions seems to be what works best for identifying fakes.

Future Works

• Expand work to classify fake human photos. Currently, the model fails to classify fake human pictures correctly.
• Implement a Stable Diffusion model, so I can have more fake image data
• Begin working on deep fake human photos instead of using the CIFAKE dataset.

References

• CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images