Deepfake Detection Challenge Solution

Deepfake Detection Challenge was about identifying media manipulated with Deepfake techniques. The main goal is to create a binary classification model that predicts whether input video is fake or not.

Data

Data represented with 4 datasets

• Training Set: dataset, containing labels for the target. The full training set is just over 470 GB. It is broken up into 50 groups ~10GB each.
• Public Validation Set: small dataset of 400 videos. This set is in use when you commit your Kaggle notebook. Accessible on the same page as Training Set.
• Public Test Set: completely withheld and is what Kaggle’s platform computes the public leaderboard against.
• Private Test Set: privately held outside of Kaggle’s platform, and is used to compute the private leaderboard.

Solution

Solution can be described with several steps.

Data preparation

• Extract 17 faces from each video with MTCNN face detector
• Save each face on disk
• Save corresponded metadata (which face from which video)
• Divide training dataset on train and validation part

Training

Train binary classificator based on Xception CNN.

Validation and inference

• Extract 17 faces from each video
• Predict label [0..1] for each face separately using pretrained model from the previous step
• Use mean value of 17 predictions to get prediction whether input video is fake or not

TODO:

• Migrate from Keras.Sequential to Tensorflow.data.Dataset (Currently NVIDIA DALI cannot work with TF Distributed Strategies, so we need to wait for the next release).
• Find working Tensorflow implementation of MTCNN and make it work with TF 2.0
• Clear data preparation and evaluation stages