A group of hackers has been caught using a government-run facial recognition system to scam companies in China.

It used the system to fake tax invoices which it then used to steal an estimated $76.2 million.

According to a report published by the Xinhua News Agency, at least two people have been charged. Prosecutors in Shanghai said the group hacked  the government-run facial recognition system by using high-definition photographs and stolen personal information.

The group bought the photographs from an online black market. A shell company that issued the fake tax invoices to China companies was also set up, a report published Wednesday said.

The Shanghai People's Procuratorate said two people identified as Zhou and Wu had been arrested. The agency said the group may have been operating the scam since 2018.

One of the suspects told authorities the group used an application to manipulate high-definition photos. This converted the photos into videos which made it seem like the faces were moving, blinking, talking and nodding.

"After obtaining the videos, we used a special mobile phone to hijack its camera. During the facial authentication process, the mobile camera would not start and the system would receive the pre-made video. The system accepted that I was in front of the camera, so I passed the certification," the unnamed suspect was quoted in saying.

Crimes involving the stealing of personal data have increased in China. Authorities are having problems keeping up with the thriving underground trade.

The Xinhua News Agency said that the cost of applications that help beat facial recognition systems is extremely low. This includes popular image manipulation apps such as Fangsong Huanlian, Remini Photo Enhancer and Huo Zhaopian.

The agency said there were now hundreds of online companies offering facial recognition hacking services with prices ranging from 30 yuan to 250 yuan ($4.5 to $38).

The government has submitted draft legislation that will impose fines of up to 50 million yuan for abusing or stealing private data.