Skip to content

Latest commit

 

History

History
27 lines (18 loc) · 1.55 KB

README.md

File metadata and controls

27 lines (18 loc) · 1.55 KB
Raw

Deepfake Detection Challenge

My solution to the Kaggle Deepfake Detection Challenge to achieve top 1% on the public and private leaderboard.

The final submission used was an ensemble of 3 models:

  • Single Frame classifier B6-EfficientNet pretrained on Imagenet
  • Single Frame classifier B6-EfficientNet pretrained on Imagenet (with Cutmix data augmentation)
  • Sequence of (5) frames classifier using an LSTM with B6-EfficientNet backbone with the cutmixed model as the pretrained weights while only training the head LSTM component of the network.
Deep Fake Detection Challenge Public Score Public LB Position Private Score Private LB Position
0.24397 12/2265 0.46099 15/2265

Example Experiments

Attempted various approaches at deep fake audio detection but decided against including them in the final submission due to not performing well during training.

See here for write up about how I approached this challenge.

References