Scientific ML: Copy-move forgery detection

The agent builds a system that classifies each image as authentic or manipulated and, for manipulated images, segments each duplicated region. It trains offline on the provided data and submits predictions through a sealed leaderboard gateway.

Copy-move manipulation, cloning one region of an image over another, is one of the most common ways scientific images get falsified, and it is hard to catch by eye precisely because the forged pixels are real pixels from the same image. Research-integrity review needs both answers: was the image altered, and where.

The metric combines the authenticity decision with an optimally matched instance-mask score over the duplicated regions. Scored submissions are quota-limited and return a public score only; the reward is the floor-normalized private score of the best-public submission.

Every run here converged on the same recipe. Each probed the test split's class balance with an all-authentic submission. Each trained a detector on self-correlation or offset-matching features, ratcheted the input resolution upward, and stacked ensembles and test-time augmentation on top.

What separated them was validation discipline. Both stronger runs reweighted their local validation to the class balance the probe had revealed; one log calls that reweighted score its leaderboard-estimating objective. The weakest run's log records no such correction. Their endgames split the same way. The weakest spent its last submissions sweeping decision thresholds across a fixed ensemble for about a point of public score. The strongest assembled a six-model ensemble and ran an oracle check showing little headroom left in its grouping stage. It stopped submitting with well over an hour unspent.

The sibling task cmfseg grades a different split against a different floor, so scores on the two pages are not comparable.