Train and integrate a CNN baseline for image detection

[ogulcanaydogan]

Background

The image detector in backend/app/detection/image/detector.py documents four methods (FFT/frequency analysis, artifact detection, metadata forensics, and a CNN classifier). The first three are implemented as statistical heuristics; the CNN classifier is marked as future ('Trained binary classifier (future)') and is not wired in.

Goal

Train (or fine-tune) a lightweight binary real-vs-AI image classifier and integrate it behind the existing detector interface, fused with the current heuristics.

Suggested approach

• Use a compact backbone exportable to ONNX so inference stays CPU-friendly and dependency-light.
• Combine the CNN score with the existing FFT/artifact/metadata signals into the final confidence, rather than replacing them.
• Keep the result explainable: report the CNN's contribution in the evidence card.

Acceptance criteria

• Inference path covered by tests with the model stubbed (CI stays offline/fast).
• Heuristic-only behavior preserved when the model is absent.
• A short doc/section on obtaining or training the model.

Self-contained and well-scoped for someone interested in vision models + provenance. Comment here to coordinate.