ArtifactNet

Inspiration To combat the growing threat of AI-generated content and ensure image authenticity in the digital age.

What it does ArtifactNet detects whether an image is real or fake and classifies it into one of three categories: human faces, animals, or vehicles.

How we built it We trained a deep learning model using EfficientNet-B2 and Focal Loss on the ArtiFact_240K dataset with heavy augmentation and validation.

Challenges Handling dataset imbalance, improving generalization to unseen deepfakes, and ensuring accurate multi-class classification.

Accomplishments Achieved 96% real/fake accuracy and 100% class accuracy, with high precision and recall across validation sets.

What we learned The power of data augmentation, test-time strategies, and architecture selection in boosting deepfake detection reliability.

What's next Deploying as a web app, expanding to video detection, and incorporating explainability features using saliency maps.

Built With

• built
• google
• matplotlib
• pandas
• pil
• pytorch
• scikit-learn
• timm
• tools:
• torchvision
• with