DeepFake Detection Framework
MSc dissertation: an Attention-Enhanced EfficientNetB7 deep learning framework for deepfake image detection, reaching 97% accuracy on 10,000+ synthetic images, with a Streamlit app for real-time prediction.
97% accuracy · 10,000+ images · Real-time prediction
Problem
Synthetic face images are now good enough to fool people, and simple classifiers overfit to the artefacts of a single generator.
Approach
- Built an Attention-Enhanced EfficientNetB7 classifier and compared it against CNN, SVM and Random Forest baselines.
- Ran statistical analysis, feature engineering, hyperparameter tuning and cross-validation on 10,000+ synthetic images.
- Shipped a Streamlit web application for real-time prediction.
Outcome
- 97% classification accuracy
- Graded 77.6% as the MSc practicum/thesis at National College of Ireland
Stack
Python · TensorFlow · EfficientNetB7 · CNN · Streamlit
Links
Related skills
Work together
Open to AI engineering roles in Dublin, across Ireland and remote, and to freelance projects. Get in touch.