polimi-ispl / icpr2020dfdc

Video Face Manipulation Detection Through Ensemble of CNNs
GNU General Public License v3.0
259 stars 100 forks source link

Regarding Top-up Training to improve the accuracy of existing deepfake model #78

Open soumyajee opened 8 months ago

soumyajee commented 8 months ago

Thanks for the work regarding deepfake detection. Unfortunately your pretrained model as well as our self trained models are not generalizing on new social media deepfakes, and are having increased number of False Positives and False Negatives. how can we further fine tune or top-up train, the existing best models on new viral deepfakes to improve detection rate. Also, when we are doing top-up training ourselves, the model weights are changing in such a way that for every video it starts predicting as Fake. How can we polish existing best models further ? Thanks