AI

Certification of Real Images through Calibrated Content Authentication

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 05:18:24 UTC.

Briefings

A 2026 study evaluates deepfake detectors against ten generators released over the last four years, finding accuracy dropped from 99.5% to 76%. Adversarial perturbations reduced all detectors below 2% accuracy, making labels unreliable. The authors propose a new detection paradigm that outputs a calibrated prediction of whether synthetic content admits a faithful reconstruction by a known generator, establishing plausible deniability for such content.

Key points

• Twenty deepfake detectors were evaluated against ten generators released between 2022 and 2026, showing accuracy decreasing from 99.5% to 76%.
• Adversarial perturbations reduced every baseline detector to below 2% accuracy, effectively inverting the detector's assigned label.

Context

Generative models synthesize high-quality inauthentic multimedia content that is already being misused at scale. This study evaluates detectors against ten generators released in the last four years to assess reliability.

Limitations

Evidence is limited to ten generators and twenty detectors; accuracy decreased over time. Adversarial perturbations break every baseline detector but do not cover arbitrary adversarial transformations. Post-hoc verifiability is eroding, with 1,116 of 3,000 Reddit images resisting reproduction by a 2022 generator but only 55 to 79 by a 2024 generator.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 18/10/2026, 02:34:56 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeibsnngbckzm2573kyypphz3dncb5p7x2qoi7wagycxw2oznifsfsy