Facia declares breakthrough deepfake detection scores

Facia has reached the point where it is scoring perfect accuracy for deepfake detection on third-party datasets, including Meta’s.

The company revealed it scored 100 percent accuracy with Meta’s Deepfake Detection Challenge (DFDC) dataset in a blog post. That dataset includes 2,100 videos manipulated with eight different techniques. Facia describes those techniques as “older,” but also suggests that they simulate “common deepfake threats.”

Facia also scored 100 percent accuracy detecting deepfakes from the FaceForensics, Celeb-DF, WildDeepfake and OneMillionFaces datasets.

A test of 51 deepfake detectors against the DFDC and Celeb-DF datasets carried out by Korean and Australian researchers earlier this year showed unimpressive detection accuracy though commercial solutions like Facia’s were not included.

The biometrics developer has also built its own deepfake dataset out of 3,430 synthetic images, using 13 generative AI tools. So far, Facia’s solution is up to 89.01 percent accuracy against this in-house dataset.

Facia explains that it uses a “multi-model ensemble” of at least 10 detection models for high precision. It has detected at least 53 attack types, with response times below a second for both cloud and on-premise deployments. In addition to a low false accept rate (FAR), Facia says it delivers the lowest false rejection rate (FRR).

Deepfake detection technology from Facia has been used to secure ID verification for businesses that require a high level of assurance and to protect remote onboarding systems for financial service providers, according to the post. The company says the same technology can also be used to detect and prevent the proliferation of non-consensual images or videos.

Facia completed an iBeta Level 2 test for compliance to the international standard for biometric presentation attack detection last year.

Article Topics

biometric dataset  |  Celeb-DF  |  deepfake detection  |  Deepfake Detection Challenge (DFDC)  |  Facia  |  Meta  |  synthetic data