Outlet Fraud Detection: Evidence-First Photo Screening
A no-label review pipeline that compares verification photos only with other photos from the same outlet.
The problem
When field teams collect many outlet photographs, manually checking every folder for recycled or off-topic evidence does not scale. A useful tool should rank suspicious photos and explain why they need review without claiming to prove fraud automatically.
How it works
The pipeline detects exact re-uploads by file hash, extracts image embeddings, and builds a per-outlet similarity view. Density, gap, minority-cluster, and duplicate rules produce review flags with deterministic reasons.
A local review interface shows the outlet image grid, rule badges, and a similarity heatmap. The repository includes a small sample outlet and committed results so the workflow can be inspected without private field data.
Results and limits
On the repository's labeled manual benchmark of 213 images from 15 outlets, the combined rules plus duplicate detection reported 83.3% precision and 85.7% recall. These are benchmark results for that sample, not a production guarantee.
The suspicion score is a rank within an outlet, not a probability of fraud. Small photo sets and missing time information limit what the tool can conclude; a human reviewer makes the final decision.