Deepfake detection

Deepfake detection was added in v1.2 and is still in experimental phase. The accuracy is amazing:

• 99.7% for “intra-testing” (Train on X and evaluate on X)

• 94.2% for “inter-testing” (Train on X and evaluate on Y)

The models are trained on a small dataset with static-images and we still perform well on most of randomly selected Youtube videos. The results are very encouraging and suggest that there is room for improvements if we train the models with videos (e.g. FaceForensics++). We’re planing to release a standalone product fine-tuned for videos.

The next video shows the result on a Youtube video from https://www.youtube.com/watch?v=2tgqX5WVhr0:

---

• To test the implementation: https://www.doubango.org/webapps/face-liveness/

• Source code used to generate the video is at https://github.com/DoubangoTelecom/FaceLivenessDetection-SDK/blob/master/samples/python/deepfake/deepfake.py.