Deepfake Detection project based on CNNs.

Getting Started

Prerequisites

• Install conda
• Environment creation :

$ conda create -n dd-cnn python=3.8
$ conda activate dd-cnn
$ pip install -r requirments.txt

• Warning: You should install TensorFlow manually to install the proper binary file for your system and then the requirements. Ex :
• ROCM: (compatibility)

# My version: 
$ pip install tensorflow-rocm==2.11.0.540

• CUDA: (tensorflow-official)

$ conda install -c conda-forge cudatoolkit=11.8.0
$ python3 -m pip install nvidia-cudnn-cu11==8.6.0.163 tensorflow==2.11.*

• Create or Download Datasets: No datasets are provided with this project, here are the prerequisites to have a usable dataset with this project:
  • Images / Frames should be in a root directory (no specified tree) and the faces must be pre-extracted, my-extraction-method using YuNet program.
  • The frames referencing must be made by a dataframe with pandas' pickle ext format.
  • dataframe mandatory content:
    • Relative path to the frames in relation to the root directory as the index of the dataframe.
    • label column with True or False values respectively for Fake and Real frames.

Available Architecture / Net :

• Xception
• EfficientNetB4
• ResNet152V2

You can easily add a new Keras model by copying the code in the models folder, for instance: Xception

Training Run Command Example on EfficientNetB4:

$ conda activate dd-cnn

# Create a save directory for models and history:
$ mkdir ./log

# Specify the necessaries path:
$ ROOT_DATASET_PATH=/path/to/your/dataset/folder
$ DF_DATASET_PATH=/path/to/your/dataframe.pkl

# Args:
$ python main.py train -h

# If using ROCM and get an error with tensorflow (current GPU is ignored) a possible solution:
$ HSA_OVERRIDE_GFX_VERSION=10.3.0

# Run command:
$ python main.py train --arch "EfficientNetB4" \
-root $ROOT_DATASET_PATH \
-df $DF_DATASET_PATH \
-o "./log" \
-s 41 \
-d "70-20-10" \
--shape 256 \
-epoch 100 \
-b 8

Plot Training History Run Command Example on EfficientNetB4

$ conda activate dd-cnn

# Args:
$ python main.py plot -h

# Path to the '.log' directory previously created in the above example
$ MODELS_LOG_DIR=/path/to/your/models/dir

$ python main.py plot -root "./log" --arch "EfficientNetB4" -m "latest"

Good to know: It does not save the graphs.

Evaluation Run Command Example on EfficientNetB4

$ conda activate dd-cnn

# Args:
$ python main.py eval -h

# Specify the necessaries path:
# Path to the '.log' directory previously created in the above example
$ MODELS_LOG_DIR=/path/to/your/models/dir
$ ROOT_DATASET_PATH=/path/to/your/dataset/folder
$ DF_DATASET_PATH=/path/to/your/dataframe.pkl
# Path to the YuNet face extraction model (This is the algorithm I used to extract the faces, feel free to change)
$ YUNET_MODEL_PATH=/path/to/your/yunet_model.onnx

$ python main.py eval --shape 256 \
-r $ROOT_DATASET_PATH \
-df $DF_DATASET_PATH \
-o $MODELS_LOG_DIR \
--arch "EfficientNetB4" \
--seed 41 \
-d "70-20-10" \
-yunet $YUNET_MODEL_PATH

Results Example on EfficientNetB4:

Credits

• Tangui Steimetz, ENSICAEN.
• Guoqiang Li, MOBAI - NTNU.