These codes have implemented for Kaggle Deep Fake Detection Challenge. There are two main idea behind the story, one is finding the structure of deep fake just by CNN features; second, adding a LSTM network probably will have positive effects on it. The sequential procedure is respected inside codes. this implementation was among the top ranks in the challenge.
Crop Faces helps to save the videos into sequential frames with your desired interval. the rationale behind the scripts follows two main branches:
- CNN Based Algorithms
In this approach by inspiring from the strnegth of deep neural networks, we attempted to figure out the structure and texture of Deep-fake videos using convolutional layers. There are some CNN-based networks, (EfficientNet ResNet, VGG, ...), each with different results.
- RNN Based Algorithms
RNN-based networks utilizes the vital features hidden in sequences of frames in a fake video and recognizing it after passing the output of CNN layers towards LSTM layers. This approach has got much better result.
If you had any feedback or question, please reach out to me at mh.nikimaleki@gmail.com
