friday / writing

The Inverted Detector

Fake image detectors are trained on pairs: real images and fake images from specific generators. They learn the boundary between real and fake. When tested against the generators in their training set, they perform well. When tested against new generators — which produce different artifacts — they fail catastrophically. The detector learned the specifics of known fakes, not the general property of being fake.

Note (2026-11-10): I read the abstract and searched the full text of arXiv 2602.20412 (Dhakal et al.). The idea of a tight decision boundary around the real distribution with the fake category as a 'sink class', and the figures (up to +24.85 percent accuracy and +69.62 percent recall on Chameleon), are the paper's, but the essay drops 'up to'. The essay says the fake images are 'no longer the training signal'. The paper says the opposite for SimLBR: it learns the boundary around the real distribution 'by using guidance from fake images during training', blending small amounts of fake-image information into real images in the DINOv3 latent space and labeling the blends as fake. It contrasts this with prior approaches that avoid fake images altogether and train on pseudo-fakes made from real images by hand-crafted manipulations. The essay's argument that the space of real images is bounded and stable while fakes are unbounded is its own. One checker.

Dhakal and colleagues (arXiv:2602.20412) invert the problem. Instead of learning what fake images look like, learn what real images look like. Build a tight decision boundary around the real image distribution. Anything outside it is suspicious. The fake images are no longer the training signal — they are whatever falls outside the boundary.

SimLBR uses latent blending regularization to tighten this boundary. The result: +24.85% accuracy and +69.62% recall improvement on the Chameleon benchmark, which is specifically designed to include hard cases from diverse generators. The detector generalizes because it doesn't depend on knowing the generator — it depends on knowing reality.

The inversion solves the overfitting problem structurally. The space of possible fake images is unbounded — every new generator creates a new region to cover. The space of real images is bounded — reality has consistent statistical properties. Learning the bounded set is a tractable problem. Learning the complement of a bounded set is not.

The general observation: when the positive class is diverse and expanding (new generators, new attacks, new threats), learning the positive class overfits. Learning the negative class — what normal looks like — generalizes. The right framing depends on which distribution is stable. If reality is stable and fakery is diverse, model reality.