Facial recognition technology using artificial intelligence has made rapid progress. Research is under way for the use of the face as a form of ID.
A group led by Professor Jun Sakuma at the University of Tsukuba's Center for Artificial Intelligence Research carried out an experiment on security involving the technology.
In the experiment, an AI learned to recognize a person's face. The team then used an adversary -- another AI -- to break into the former and create data of faces similar to the one it recognized.
The adversary showed the faces to the first AI. Based on how the first AI reacted, it came up with closer approximations. The learning process was repeated at high speed.
In about 2 days, the adversary was able to create a facial image that the first AI recognized as the real face.
Researchers say immediate risks of abuse may not be high because it is difficult for an adversary to read the AI program of its target system under normal circumstances.
But the team still cautions against potential risks of a hacker recreating faces and tricking facial recognition.
Sakuma says developers of AI did not prepare for the risks of stored data being accessed by a third party. He says such data may include business secrets or individual faces. He calls for steps to deal with risks of information leaks.
Wider applications of facial recognition
Facial recognition is finding its way into daily life.
Apple's latest smartphone model, iPhone X, uses the technology to allow its owner to unlock the phone.
A US hamburger chain is also experimenting with the technology. It suggests menu items based on past purchases. The system also uses facial recognition to authorize payments.
Besides facial recognition, AI-based image recognition is expected to be used in various fields. The technology can be used to help self-driving vehicles detect a dangerous object. It can also be used in medical image processing to assist diagnosis.
Professor Sakuma says if an attacker could hack an AI, the owner's facial data and other personal information could leak.
Besides data breaches, Sakuma warns that overriding facial recognition would allow a person to use a computer or other device freely. He says it's important to prepare for the worst.
Expert: Problems must be addressed
Besides security, another issue with AI-based image recognition is errors.
A team at the Massachusetts Institute of Technology created a model of a turtle and showed it to an AI. It says the system recognized it as a rifle.
Researchers at the University of Washington carried out an experiment in which they created a stop sign and put a sticker on it. They say an AI misidentified it as a speed-limit sign.
An AI expert at the University of Tokyo, Yutaka Matsuo, says there has not been much discussion on the vulnerabilities of AI. He says AI is still developing, and it's necessary to deal with security and other challenges.