“This work represents a major step ahead in strengthening our info benefit as we fight refined disinformation campaigns and synthetic-media threats,” says Bustamante. Hive was chosen out of a pool of 36 corporations to check its deepfake detection and attribution know-how with the DOD. The contract may allow the division to detect and counter AI deception at scale.
Defending towards deepfakes is “existential,” says Kevin Guo, Hive AI’s CEO. “That is the evolution of cyberwarfare.”
Hive’s know-how has been skilled on a considerable amount of content material, some AI-generated and a few not. It picks up on indicators and patterns in AI-generated content material which are invisible to the human eye however could be detected by an AI mannequin.
“Seems that each picture generated by one in every of these mills has that form of sample in there if you realize the place to search for it,” says Guo. The Hive staff always retains monitor of latest fashions and updates its know-how accordingly.
The instruments and methodologies developed via this initiative have the potential to be tailored for broader use, not solely addressing defense-specific challenges but in addition safeguarding civilian establishments towards disinformation, fraud, and deception, the DOD stated in an announcement.
Hive’s know-how gives state-of-the-art efficiency in detecting AI-generated content material, says Siwei Lyu, a professor of pc science and engineering on the College at Buffalo. He was not concerned in Hive’s work however has examined its detection instruments.
Ben Zhao, a professor on the College of Chicago, who has additionally independently evaluated Hive AI’s deepfake know-how, agrees however factors out that it’s removed from foolproof.
“Hive is actually higher than a lot of the industrial entities and among the analysis strategies that we tried, however we additionally confirmed that it isn’t in any respect arduous to bypass,” Zhao says. The staff discovered that adversaries may tamper with photos in a approach that bypassed Hive’s detection.