Ů��С��Ƶ

The new iPhone X puts face recognition front and centre. Why? Because it is the quickest and easiest way to unlock your phone. Before you’ve even thought about placing your finger on the scanner or entering a PIN, the camera will have already processed your face, checked it’s you, and granted access.

Despite its handiness, however, the technology is far from proven. Apple keeps its algorithms under lock and key, meaning we don’t know for sure how well they work, but even the best in the business have woefully low reliability. “I wouldn’t recommend anyone trust an iPhone’s face verification,” says at Imperial College London.

And Apple isn’t the only one using software to recognise faces. Everyone from border control to high-street shops to police forces is trying to get in on the action. How much can we actually trust the technology? And do we want to live in a society where your face is always being tracked?

The best current face-recognition algorithms use a form of artificial intelligence called deep learning. Companies trawl the web to gather billions of images and use them to train an algorithm inspired by neurons in the brain, called a deep neural network. Slowly, the algorithm learns to extract features from an image that are relevant to your identity, like the position of your nose or the gap between your eyes. The more images it sees, the better it becomes.

When your face appears in front of a camera, this network kicks into action. The iPhone X also has another trick: a 3D camera that means if your head is tilted away from the screen, the algorithm can work out what it would look like if you were facing straight at it. Then if enough features match, bingo! You are successfully verified and your phone is unlocked. If not, access is denied.

“The chance that a random person in the population could unlock your iPhone X and unlock it with their face is one in a million,” said Apple’s Philip Schiller during a press conference on 12 September.

But in the real world…

However, Apple are only talking about one type of error that could happen with its Face ID algorithm – someone else gaining access. But what if your phone refuses to recognise your face, and doesn’t let you in? That’s a problem too. “The two errors are related; for a given biometric system, if you reduce one type of error, the other type of error goes up,” says at Michigan State University.

Under normal conditions, such as without perfect lighting, a perfectly still camera and the person in the perfect pose – in other words, the real world in which we use our phones, publicly tested algorithms have only managed to get below an overall . This is a high risk of your phone failing to unlock, or, worse, unlocking for someone else.

When demonstrating Face ID on stage at the press conference, Apple’s Craig Federighi found this out for himself. The software failed to recognise him and he ended up having to enter a passcode to unlock the phone.

“We have to wait till November [when the phone is released] to see the usability based on people’s experience when they get the phone,” says Jain.

Until then, as Apple hasn’t demonstrated the resilience publicly, it seems silly to put too much faith in their software. And while you might be comfortable with using your face as a password, increasingly there is more at stake than a few text messages.

Family pictures, personal emails, those salacious WhatsApp messages, and even banking apps are all vulnerable once your phone is unlocked. Apps like Apple Pay and Android Pay allow you to flash your phone instead of your credit card, so theoretically anyone with access to your unlocked iPhone can spend an unlimited amount, although Android currently set a limit of £30 in the UK per transaction. The iPhone X will allow users to use Face ID to authorise transactions instead of PIN, and Android is expected to follow suit with something similar soon.

Why the long face?

Face recognition on its own may not be secure enough to protect our digital vault, but that doesn’t mean we should ditch it altogether. Instead, we can combine it with other biometrics. Fingerprints, iris scans and even voice recognition could all help provide a speedy authentication method.

There is an air of inevitability in the payments world that something about you will be your password in the future, rather than something you have to remember, says David Baker at UK Finance, the trade association for the UK banking sector. Although there is still a long way to surpass PIN on ease of use, he adds.

Faces aren’t just taking off on the iPhone. What most people know as “face recognition” is more properly called “face verification”, the task of confirming that one person’s face matches their records. True face recognition involves matching a face to a vast database of faces, like police trawling CCTV images for potential suspects. Although this is a much harder task, lower degrees of accuracy are required, because a human police officer will make the decision to further investigate a match.

This is already happening in the UK, where happened just last year, but the tech is still pretty basic. Police have been trialling the software at various large events, such as the Champions League Final and Notting Hill Carnival. At this year’s carnival the system flagged , so there’s still some way to go. Currently in the UK, the software is only used with bespoke cameras mounted on top of vans, but the same tools could be integrated with CCTV that is already installed.

Lack of transparency

Even though the technology is still in its infancy, the databases used alongside it are not. In England and Wales, the police have amassed a massive database of more than 20 million facial images. Many of these are mugshots, but many are also from interviews with innocent people who have been questioned during a case. In 2012, the High Court said it was , but the government has said the police do not need to delete them unless they receive specific requests from the individuals concerned.

Things are no better in the US. Approximately half of US adults appear in a face database amassed by the FBI, with about 80 per cent of the photos from people who haven’t been connected with a crime. They include driving license and passport pictures.

These huge databases are worrying, but what is that more worrying is we don’t fully know how they are being used. The software isn’t accurate enough to reliably issue arrest warrants or fines without a human checking the match first, but individual officers may not know that, and could be influenced by unconscious bias when presented with an “infallible” machine’s results.

“People thought deep learning would solve everything. But companies like Google have trained their algorithms on billions of images and they are still not very good in arbitrary conditions,” says Zafeiriou.

Not that improving the algorithms would solve this thorny issue. If CCTV cameras could identify every person who passes it, the temptation to track everyone all of the time might just be too great for states to resist. So when you get in line for your new iPhone, think carefully about the world you are signing up for.