TODO.md

Proof of Concept

• Add RESNET50 model to proof of concept
• Start write up for main diss
• Make demos (live view, noise visualisation, etc.)

Main model

• Research and implement custom model for eye landmark detection (cropping on eye, then some kind of pre-trained model?)
  • Model is from https://www.sciencedirect.com/science/article/pii/S0031320319303772
  • Create and annotate dataset (email authors?)
  • Train model
  • Test model
• Implement EAR analysis
  • Could do a 1d CNN
  • split each blink into a feature vector (length of down, length of up, period closed, period between blinks, etc.)
  • non-ML-based analysis (thresholds, entropy, etc.)
  • or a combination of the above

Noise

• Code own versions of all the noise functions
  • CW-L2
  • FGSM
  • FakeRetouch

Testing

• Email SCRTP to ask for more space
• See if anyone has got access to Meta's DFDC or try and find an email
• Look into complete deepfaked models (not changing the face with another clip, but changing the face with a generated face)
• Adapt script to be generalised (python test.py <path_to_dataset>)
• Work on test script to make multithreaded and save progress as it goes (add buffer back in to speed up testing)

Presentation

• Make presentation
• Practice presentation
• Present presentation (ideally will have final data by this point)

Dissertation

• Words (a lot of them (an awful lot of them))