Facebook AI Research has unveiled a new de-identification engine that uses machine learning to remove identifying facial information from video content. The platform combines an adversarial autoencoder with a classifier network to create two images of a person’s face – one distorted and one not – that can then be embedded in a video for distribution.
In plain terms, that means that while human viewers will still be able to recognize the people in the video, those individuals will not generate hits when the video is fed to a facial recognition algorithm. The process will ensure the privacy of the subjects without needing to scrub their faces entirely.
Unlike de-identification solutions that will only work with still images, Facebook’s tech can be applied to live and recorded video. The company also claims that the same techniques can be used to mask other biometric identifiers, including voice and behavioral characteristics.
At the moment, Facebook does not have any plans to integrate the system into its consumer offerings, although the technology could represent a potential solution to the company’s ongoing privacy issues. Facebook is currently embroiled in a class-action lawsuit that accuses the company of using facial recognition without receiving sufficient consent, and has since given users the option to opt out of facial recognition on the social media network.
Of course, Facebook is not the only company working to develop de-identification technology. D-ID recently released a Smart Anonymization platform that builds on its previous facial recognition jammer and allows clients to remove Personally Identifying Information from photo and video materials.
Source: Venture Beat
October 28, 2019 – by Eric Weiss