According to estimates by Digtimes Research, the smartphones of Chinese manufacturers will deliver 199.1 million…
Filters are designed to protect people’s photos from face detection
Researchers at the University of Toronto have developed an algorithm to dynamically interrupt facial recognition systems. In the form of a filter, users of Instagram, Facebook and Co can protect their photos from face analysis in the future. The new approach uses a learning method called adversarial training, in which two artificial intelligence algorithms compete against each other.
“Privacy has become a real problem as facial recognition is getting better,” explains researcher Parham Aarabi. “This is a way in which beneficial anti-face recognition systems can combat this ability.”
Filter changes pixels
Aarabi and colleague Avishek Bose have therefore designed two neural networks: one that can identify faces and a second that wants to disrupt the face recognition task of the first one. The two fight and learn from each other constantly and organize a continuous AI arms race. The result of the data obtained is a type of filter that can be applied to photos to protect the privacy of the depicted user. The algorithm alters very specific pixels in the image and causes such small changes that they are almost imperceptible to the human eye.
“We create very fine perturbations in the photo, but they are significant enough for the detector to fool the whole system,” says Bose.
In addition to disabling facial recognition, the filter also interrupts image-based search, feature recognition, emotion and ethnicity assessment, and all other facial attributes that can be automatically extracted.
In the next step, the researchers want to make the filter accessible to users. This should happen either through a smartphone app or through a website. Especially for social networks like Facebook or Instagram would be worth a private application. Since the tiny deviations hardly noticeable, photos are not disturbed in their appearance.