|
--- |
|
license: apache-2.0 |
|
metrics: |
|
- accuracy |
|
- f1 |
|
base_model: |
|
- google/vit-base-patch16-224-in21k |
|
--- |
|
Checks whether an image is real or fake (AI-generated). |
|
|
|
**Note to users who want to use this model in production** |
|
|
|
Beware that this model is trained on a dataset collected about 3 years ago. |
|
Since then, there is a remarkable progress in generating deepfake images with common AI tools, resulting in a significant concept drift. |
|
To mitigate that, I urge you to retrain the model using the latest available labeled data. |
|
As a quick-fix approach, simple reducing the threshold (say from default 0.5 to 0.1 or even 0.01) of labelling image as a fake may suffice. |
|
However, you will do that at your own risk, and retraining the model is the better way of handling the concept drift. |
|
|
|
See https://www.kaggle.com/code/dima806/deepfake-vs-real-faces-detection-vit for more details. |
|
|
|
``` |
|
Classification report: |
|
|
|
precision recall f1-score support |
|
|
|
Real 0.9921 0.9933 0.9927 38080 |
|
Fake 0.9933 0.9921 0.9927 38081 |
|
|
|
accuracy 0.9927 76161 |
|
macro avg 0.9927 0.9927 0.9927 76161 |
|
weighted avg 0.9927 0.9927 0.9927 76161 |
|
``` |
