Deepfake Duel: Truth vs. Trickery_code crafters

• 

  CNN using Transfer Learning

Note

We have attached the link of Google Drive for model weights. Check LOOM for our video demo

Inspiration

The day we started learning Computer vision during one of our courses, me and Nandish, we were fond of Image classification problems. This was one of the toughest and most time-consuming challenges.

What it does

We trained two models: one for classifying the classes into Human faces, animals, and vehicles and the other model we trained was for classifying real vs fake.

How we built it

We used the concept of Transfer Learning. VGG16 and Xception model was used.

Challenges we ran into

Huge dataset and less computational resource on Google Collab, which gets deactivated after training for 3-4 hours. Unable to run on kaggle.

Accomplishments that we're proud of

We classified the images into human_faces, animals, and vehicles with an accuracy of 99.93% and deep vs fake with 71.55%, which is still better than randomly guessing, that is 50%.

What we learned

It was really difficult to distinguish between real and fake images even with naked eyes. Still the model got an accurcy of 75% was astonishing. This shows the power of transfer learning and CNN architecture.

What's next for Deepfake Duel: Truth vs. Trickery_code crafters

Due to time constraints and less computational powers we were not able to look at more different architectrures

Built With

• cnn
• dataaugmentation
• deeplearning
• python
• transferlearning
• vgg16
• xception