Certification of Real Images through Calibrated Content Authentication
The paper argues that “real image” certification should hinge on whether known generators can faithfully reconstruct the image, not on detector confidence alone.
Its tests found deepfake detector accuracy falling from 99.5% to 76% across newer generators, and adversarial perturbations pushed every baseline below 2% accuracy. The authors frame this as a provenance problem: image content alone may not prove authenticity if generators can reproduce authentic-looking content. Their proposed calibrated system treats successful reconstruction by a known generator as plausible synthetic provenance, while bounding how often generated images are wrongly certified. In one reported setting, calibration limited wrongly certified generated content to at most 1%, though the paper says this does not cover arbitrary adversarial transformations. HF Daily Papers' note
Its tests found deepfake detector accuracy falling from 99.5% to 76% across newer generators, and adversarial perturbations pushed every baseline below 2% accuracy. The authors frame this as a provenance problem: image content alone may not prove authenticity if generators can reproduce authentic-looking content. Their proposed calibrated system treats successful reconstruction by a known generator as plausible synthetic provenance, while bounding how often generated images are wrongly certified. In one reported setting, calibration limited wrongly certified generated content to at most 1%, though the paper says this does not cover arbitrary adversarial transformations. HF Daily Papers' note
score 5