The most popular and comprehensive Open Source ECM platform
Let us know more
Artificial Intelligence: Facial Recognition Vulnerable to Attacks and Misuse
How secure is facial recognition as a biometric tool to lock up secrets? It may be good enough to protect access to your cellphone, but increasingly adversarial AI researchers are figuring out ways to defeat it.
Researchers at Tel Aviv University in Israel developed a set of master faces that can act as a skeleton key capable of impersonating somewhere between 40-60 percent of human faces.
The researchers wrote that “our results imply that face-based authentication is extremely vulnerable, even if there is no information on the target identity. In order to provide a more secure solution for face recognition systems, anti-spoofing methods are usually applied. Our method might be combined with additional existing methods to bypass such defenses.”
The researchers generated fake images of faces are that contain markers which match facial features common to large numbers of people, creating an “omni-face”. Next the fake images can be paired with deepfake technology to animate the face so as to avoid AI tests that determine whether the image is a living person or not. The researchers created nine of these faces and were able to break facial recognition protections more than 40 percent of the time using just one of these images. The researchers are calling these images “Master Faces”.
The researchers said that “generating more faces would likely cover an even broader range of the population, making the attack yet more foolproof – an even more worrying development.”
The result of this study adds more weight to the argument against using facial recognition technology. Opponents of the technology have long cited concern with racial profiling and mass surveillance. False-positive inaccurate matches have also occurred frequently, especially for images of Black subjects. The UN has proposed banning state use of facial recognition for some time.
