Israeli company Adversa has developed the Adversarial Octopus technology to fool facial-recognition algorithms into identifying one person as another. Motherboard reports.
To demonstrate its capabilities, the developers edited the photo of the company’s chief executive, Alex Polyakov, and uploaded it to PimEyes, a public face-recognition service. The service misidentified him as Elon Musk.
Unlike similar technologies, Adversarial Octopus is more adaptable, stealthy and accurate, Polyakov said.
\”Other methods simply hide you without swapping you for someone else,\” he added.
According to Polyakov, Adversarial Octopus can undermine the training of facial-recognition systems by adding noise to images from the training dataset. The method does not require internal knowledge of how the system was trained.
He also acknowledged that the algorithm is a ‘black box’ and its creators do not understand the logic by which neural networks achieve their objective.
The company has not yet published a peer-reviewed study explaining Adversarial Octopus. Polyakov said they plan to publish data after the responsible disclosure process: informing firms involved in facial recognition about vulnerabilities and ways to defend against them.
In May 2021, researchers introduced two free tools forprotecting portrait photographs from facial-recognition systems — Fawkes and LowKey.
In April, the DoNotPay startup developed a service forprotecting images from biometric identification by face.
Subscribe to ForkLog AI on Telegram: ForkLog AI — all the news from the world of AI!