CoFi-E: Repairing Blind Spots in Conformal Detectors
Published:
- Identifies rank-fiber collisions: distribution changes that leave a detector’s score distribution unchanged, so every conformal bettor on that score is blind while false-alarm control still holds.
- Refines the score by splitting groups of observations that share a score whenever training alternatives reveal a difference inside them; independent audit data certify each accepted split.
- Refinements never discard rank information the detector already exposes, and hidden information shrinks geometrically when informative splits are available.
- Runtime false-alarm control is kept separate from learning: the refined score is frozen before fresh reference data are drawn.
- Evaluated on MVTec AD, UCR, and UEA benchmarks, with transfer bounds that quantify how detection evidence depends on an unseen anomaly family’s distance from the training alternatives.