Google claims its facial recognition system can achieve near 100 percent accuracy

Google published a paper detailing its new artificial intelligence system, FaceNet, which Google researchers call the most-accurate technology available for recognizing human faces, according to a report by Fortune.

In the paper, titled “FaceNet: A Unified Embedding for Face Recognition and Clustering”, Google claims the system achieved nearly 100-percent accuracy rate on the facial recognition dataset Labeled Faces in the Wild.

The dataset uses more than 13,000 face images from the Internet to measure how accurate the algorithms are at verifying whether two images are of the same individual.

FaceNet was trained on an enormous 260-million-image dataset and performed at an 86 percent and higher accuracy rate.

Last year, a team of Chinese researchers claimed to have achieved better than 99 percent accuracy.

In the June 2014 paper, Facebook researchers said that humans analyzing images in the Labeled Faces dataset are only able to achieve 97.5 percent accuracy.

However, Google researchers went beyond simply verifying whether two faces are identical by also attaching a name to a face.

Taking a classic facial recognition approach, the researchers went as far as organizing the images into groups of faces that look the most similar or the most distinct.

When Google’s FaceNet and Facebook’s “DeepFace” are eventually integrated into their company’s web platforms, they will automate the process of tagging photos and searching for people and will also make it easier for web companies to analyze their users’ social networks and to assess global trends.

Article Topics

 |   |   |   |   | 

Comments

18 Replies to “Google claims its facial recognition system can achieve near 100 percent accuracy”

  1. For convenience or for security?

    Near-100% accuracy still involves the trade-off between false rejection and false acceptance, which means that system requires a fallback password in case of false rejection.

    For biometrics to displace the password for security, it must stop relying on a password registered in case of false rejection. Threats that can be thwarted by biometric products operated together with fallback/backup passwords can be thwarted more securely by passwords only.

    We could be certain that biometrics would help for better security only when it is operated together with another factor by AND/Conjunction (we need to go through both of the two), not when operated with another factor by OR/Disjunction (we need only to go through either one of the two) as in the cases of biometric products on the market that require a backup/fallback password, which only increase the convenience by bringing down the security.

    Biometric solutions could be recommended to the people who want convenience rather than security but should not be recommended to those who want security rather than convenience.

Leave a Reply

Brand Focus

Biometrics Research Group

Biometrics White Papers

Biometrics Events

Explaining Biometrics