NewThe detectors that scored perfect collapsed the hardest under attack.
Margen · ResearchFindings first

Open, cited research

The lab.

We publish open benchmarks of the detectors meant to catch deepfakes, measured against the generation methods adversaries actually use and the conditions media really arrives in. The point is simple: show, in public, what detection can and cannot do, so buyers and vendors do not have to take anyone's word for it.

14
detectors benchmarked
12
demographic groups
100%
open methodology

Published in the open, with a permanent DOI you can cite. View the citation

Featured
BenchmarksPublished paper

Detectors collapse from near-perfect to near-random.

An initial benchmark and robustness re-evaluation of leading open-source deepfake detectors, stratified across generators, platform degradations, and demographic groups.

Open data · permanent DOI

6 min readRead the analysis
Research partnerships

Building a benchmark together?

We collaborate with research groups, standards bodies, and competitions on open benchmark design, evaluation, and sponsorship. If you work on how to measure synthetic media, we would like to talk.

Partner on research

Field notes

Quarterly changes in attacker tooling, in your inbox.

A short briefing on shifts in attacker tooling, four times a year.

Research and benchmarks only. Unsubscribe anytime.