Project aims to achieve an accuracy rate of 90 per cent but faces formidable technological hurdles and concerns about security
South China Morning Post
The goal is for the system to able to match someone’s face to their ID photo with about 90 per cent accuracy.
The project, launched by the Ministry of Public Security in 2015, is under development in conjunction with a security company based in Shanghai.
The system can be connected to surveillance camera networks and will use cloud facilities to connect with data storage and processing centres distributed across the country, according to people familiar with the project.
However, some researchers said it was unclear when the system would be completed, as the development was encountering many difficulties due to the technical limits of facial recognition technology and the large population base.
At present, similar systems operate on a smaller level, including police databases and city or provincial ID pools.
But these operate separately and are on a much smaller scale.
There is also a national database of police suspects and people of interest to the government.
These may continue to be used independently after the national system is established.
The core data set for the national system, containing the portrait information of each Chinese citizen, amounts to 13 terabytes.
The size of the full database with detailed personal information does not exceed 90 terabytes, according to technical documents on the ministry’s website and a paper written by police researchers.
Chen Jiansheng, an associate professor at the department of electrical engineering at Tsinghua University and a member of the ministry’s Committee of Standardisation overseeing technical developments in police forces, said the system would have to be built on an unprecedented scale because no country had a population as big as China’s.
The system was being developed for security and government uses such as tracking wanted suspects and public administration, he said.
Commercial application using information sourced from the database will not be allowed under current regulations.
“[But] a policy can change due to the development of the economy and increasing demand from society,” Chen said.
Giving commercial sectors access to the database under proper regulation would create new business opportunities by helping to improve customer service, he said.
Chinese companies are already taking the commercial application of facial recognition technology to new heights.
With a smile or blink of the eyes to a camera, students can now enter their university halls, travellers can board planes without using a boarding pass and diners can pay for a meal at KFC.
Some other restaurants have even offered discounts to customers based on a machine that ranks their looks according to an algorithm. Customers with “beautiful” characteristics – such as symmetrical features – get better scores than those with noses that are “too big” or “too small” and those that get better scores will get cheaper meals.
Some public lavatories in Beijing also use facial recognition so that the automatic dispensing machines will deny toilet paper to people who ask for it more than once within a given period.
Facial recognition could supersede other personal identification methods that are used to make payments such as scanning fingerprints or QR codes on a mobile phone.
But the government project has prompted controversy among artificial intelligence experts.
Cheng Mingming, a professor of computer science at Nankai University in Tianjin, said that despite the scale of the project, technological advances meant that all the information could be stored in small, portable drives – which raised the risk of data theft.
He said a palm-sized commercial hard drive nowadays could store 10 terabytes or more of data and you could “pack it in a suitcase and board a flight”.
“If the facial data and related personal information is stolen and put on the internet, it will cause big problems,” Cheng said.
For instance, due to the rapid advance of facial recognition technology, a person or organisation could take a photo and identify strangers at a party or on the street without their knowledge, Cheng said.
But a network security vendor for the Ministry of Public Security dismissed the possibility.
“To download the whole data set is as difficult as launching a missile with a nuclear warhead. It requires several high-ranking officials to insert and turn their keys at the same time,” the vendor said.
The 1.3 billion-person facial recognition system is being developed by Isvision, a security company based in Shanghai.
Isvision confirmed to the South China Morning Post that it had won the contract last year but declined to provide details.
“The progress of development is confidential. At present we have no information for public disclosure,” a company spokesperson said.
Isvision security cameras with facial recognition capabilities were first deployed in Tiananmen Square as early as 2003, according to the company’s website.
The system was connected to the police database of suspects, capable of recognising and tracking potential targets in a large crowd.
The company has also set up similar systems for law enforcement authorities in Xinjiang and Tibet, where riots have broken out from time to time because of serious ethnic conflicts.
According to Fan Ying, a researcher at the ministry’s population management research centre in Beijing, the project team has encountered “unprecedented challenges” due to the government’s high demands for speed and accuracy.
When a photo, gender and age range are inputted, the system is required to find a match within three seconds with an accuracy level higher than 88 per cent.
Fan and colleagues tested the facial recognition algorithm developed by Tsinghua University, a world-leading institute in this field of research, and they were disappointed with the results.
They found that the accuracy of the photo that most closely matched the face being searched for was below 60 per cent. With the top 20 matches the accuracy rate remained below 70 per cent, Fan and collaborators reported in a paper published in the domestic journal Electronic Science and Technology in May.
“It cannot solve problems with real-life applications,” they added.
The system developed by Isvision will use an algorithm developed by SeetaTech, a start-up established by several researchers from the Institute of Computing Technology at the Chinese Academy of Sciences in Beijing.
SeetaTech confirmed to the SCMP its involvement in the national facial recognition project but declined to comment further.
A researcher at the Institute of Computing Technology familiar with the project said some huge technical hurdles remained.
“Among 1.3 billion people, some totally unrelated people have faces so alike even their parents cannot tell them apart,” the researcher said.
“Currently the access to the database is limited to a few security companies with very close ties with the Ministry of Public Security.
“More access will definitely lead to higher risk of [data] leakage.”
The researcher warned that the cost of the convenience facial recognition could bring to everyday life was “sacrificing security”.