Neural network learns to fool identification systems with a ‘universal’ face

Researchers at Tel Aviv University in Israel developed a method to bypass biometric identification systems using a forged facial image.

The disclosed method uses artificial intelligence technologies to create a universal facial template that can sequentially merge with and unlock identity verification systems.

According to the researchers, the vulnerability lies in the fact that facial recognition uses a wide set of markers to identify specific individuals. With these, a universal template can be created that can fool a large percentage of security systems.

Researchers suggested that as few as nine faces generated by the algorithm StyleGAN could correspond to 40% of the Earth’s population. They tested the neural-network-synthesized image on a dataset of 13,000 images and found that the fake face could imitate 20% of the identities in the dataset. Other tests showed higher results.

“Face authentication is extremely vulnerable, even if attackers have no information about the target identity,” the researchers said.

The researchers also believe that the vulnerability could be combined with deepfakes to mimic a “liveness” check, which is often used for remote identity verification by performing a series of gestures.

Earlier in July, researchers reported the discovery of a method for covertly inserting malicious code into a neural network.

In May, researchers from the Maryland Center for Cybersecurity identified a vulnerability in neural networks that increases their energy consumption.

Subscribe to ForkLog’s news on Telegram: ForkLog AI — all the news from the world of AI!