As part of the thesis, we set ourselves the task of creating a model that would effectively identify deepfakes.
We entered the DeepFake Game Competition, where we recognized the fakes submitted by our teammates.
Through the process of developing models for the detection of deepfakes, we tested different models and ideas.
We divided images of faces into several smaller ones.
Ones that would cover the part of the face which we assumed would become distorted when the fake was created.
During the competition, we were limited by the requirements set by the organizers.
To make the results comparable, we taught the models on a comparable set of data through the whole process, even after the end of the competition.
After the publication of the databases generated by the competitors on which our models were tested, we found that they contained a high amount of noise.
We focused on improving the result on the competition dataset and added an autoencoder to the pipeline that would detect the noise.
The final model consists of the Xception, EfficientNet, and Skip-GANomaly models learned from 2-D second level wavelet transformed input images.
With this model, on the mentioned dataset, we achieve the area under the ROC curve equal to 0.645902, which would place us in 11th place out of 28 competitors in the competition.
|
