All projects
PrototypeArtificial Intelligencevisibility: sanitizedIndependent project
AI-Generated Image Detection Prototype
PyTorch prototype combining an EfficientNet backbone with frequency-domain features
Code prototype. No evaluation results are available, so no detection accuracy is claimed.
- Role
- Developer
- Links
- Source is private or sanitized; no public repository is linked.
- Reading mode
Project overview
A code prototype exploring whether spatial features from an EfficientNet backbone combined with frequency-domain statistics (FFT magnitude spectra) can separate AI-generated images from photographs. The architecture, data pipeline and training loop were written; a rigorous evaluation on a held-out dataset has not been completed.
Problem and context
Generated images often leave spectral artifacts that spatial classifiers alone may miss; a dual-branch model could be more robust.
Objectives
- Build a dual-branch model
- Create a reproducible data pipeline
- Establish an evaluation protocol
My contribution
- Model definition, feature extraction and training loop
Technology stack
- Python
- PyTorch
- torchvision
- EfficientNet
- NumPy
- FFT analysis
Results
- Code prototype runs end to end on sample data
Limitations
- No measured accuracy; no detection performance claims
- Generalization to new generators untested