Deepfake detection made with the big data engine Apache Spark and the Neural Network library Keras.
PySpark is used to distribute the preprocessing of a labelled dataset containing >3000 videos, while Keras is used to build and train a Convolutional Neural Network to classify videos as real or fake.
The project is done as a part of the module "CS4225/CS5425 Big Data Systems for Data Science" at the National University of Singapore.
The trained model can be used on a local video or a YouTube video with a command-line interface. This will output the classification of real or fake, and how confident the decision is.
python predict.py --url <youtube-video-url>
python predict.py --path <path-to-video>
- OpenCV:
pip install opencv-python - CVLib:
pip install cvlib - Pyspark:
pip install pyspark(and some other configurations) - Numpy:
pip install numpy - Keras:
pip install keras - Tensorflow:
pip install tensorflow - MatPlotLib:
pip install matplotlib - Pydot:
pip install pydot - Pafy:
pip install pafy - YouTube-DL:
pip install youtube_dl
Or simply:
pip install opencv-python cvlib pyspark numpy keras tensorflow matplotlib pytdot youtube_dl pafy
The dataset should be stored in a folder named data, using the dataset gathered with these instructions: https://github.com/ondyari/FaceForensics/blob/master/dataset/README.md
The preprocessing is now done with the assumption that the following commands are run:
python download-FaceForensics.py path/to/project/data -d DeepFakeDetection -c c23 -t videos
and
python download-FaceForensics.py path/to/project/data -d DeepFakeDetection_original -c c23 -t videos
The preprocessing is distributed with PySpark. It can be run locally, standalone or in a cluster.
The following command can be used when in standalone mode (after setting up master and workers):
spark-submit
--conf spark.driver.memory=8g
--conf spark.executor.memory=8g
--conf spark.memory.offHeap.enabled=true
--conf spark.memory.offHeap.size=8g
--conf spark.driver.maxResultSize=8g
--master local[*] spark_preprocess.py
This runs the preprocessing with the default parameters, and stores the result in two numpy arrays; X and y. X contains samples faces and y contains their corresponding label of real or fake.
The real videos will have a higher amount of samples in order to balance the labels.
The model is a CNN, and can be trained in CNN.py or the notebook (with plotting) train_model.ipynb
It has the following architecture:
