Certification of Real Images through Calibrated Content Authentication

Sarim Hashmi, Abdelrahman Elsayed, Mohammed Talha Alam, Samuele Poppi, Nils Lukas

Generative models can synthesize high-quality inauthentic multimedia content that is already being misused at scale. We evaluate twenty deepfake detectors against ten generators released in the last four years and find accuracy decreasing over time, from near-perfect 99.5% to 76%. Adversarial perturbations further reduce every baseline detector to below 2% accuracy, effectively inverting the detector's assigned label. We argue that this unreliability reflects a fundamental ambiguity: generators can reproduce authentic content exactly (e.g., through memorization), so content alone cannot reveal the true provenance label. For this reason, content produced by a generator must admit a faithful reconstruction by that same generator, and finding such a reconstruction makes synthetic provenance plausible and authenticity plausibly deniable. We therefore propose and evaluate a detection paradigm that outputs a calibrated prediction of whether authenticity is plausibly deniable: a faithful reconstruction by any known generator establishes plausible deniability, while calibration bounds how often content from known generators fails to be reproduced. Our evaluation shows that (i) our…