Detecting deepfakes has become an increasingly important issue. This fake-detector repo contains three applications which are currently stand alone but may become integrated as part of future work.
They combine to produce a web app that predicts if an uploaded image is a deepfake and provides feedback to the user what regions of the image the model found “most important” for its prediction.
This explainability (i.e. explainable AI) helps users learn to spot deepfakes more easily without the aid of technology.
Instructions:
- Upload a headshot of a person (torso + head) in jpg or jpeg format
- Review prediction and probability (how sure the model is)
- Review the four images printed below the originally uploaded image. The overlaid heatmap (green hot, red cold) shows where the model had the highest activation values.