SoftiDoc builds biometrics for South African faces, documents and fraud patterns. Its software quietly powers onboarding for credit bureaus, retailers and micro-lenders, using a selfie and an ID photo to confirm in seconds that the person on screen is the person on the card.
“It's computer vision at the end of the day, looking at things as a computer would and mimicking what a person would do,” explains Jeremy Engelbrecht, founder and MD of SoftiDoc. The company sits in the middle of a client's process rather than in front of the customer, plugging into the pipeline at the point where an application or a login begins.
Reading a real face
An applicant receives an SMS with a link, opens a small web app and is asked to photograph their ID and then take a selfie. Within those few seconds, SoftiDoc’s system establishes that a live person is holding the camera before it compares that face to the picture on the document.
“Your nose is further than your cheekbones, your lips and obviously your hair, so we take depth maps,” says Engelbrecht. “Then there's also texture analysis, looking at how the light reflects off your skin. Those are the technologies we use to see if you're a real person, and we can do that from a smartphone.” Smartphone resolution is what made the move off specialist hardware possible, because a sharper image shows more of the face.
Spotting the spoof
Engelbrecht says there are three ways to fool a liveness check: a replay, where a face is recorded on one phone and presented to another; a printed cut-out held flat in front of the lens; and a silicone mask worn over the whole head. "We generally look at the context of the whole image, not just the face," he says.
Light behaves differently on silicone than on skin carrying its own oils, a flat image has no depth, and the pixels around a mask never blend into the background. SoftiDoc uses a Fourier transform to read signals across the whole frame, and no verdict rests on one algorithm. About seven models each check a different property and average their outputs against a threshold, trained on more than one million real faces, replays and masks. All of it is tested against iBeta, the international lab that certifies spoof detection.
Local by design
SoftiDoc is, according to Engelbrecht, the only local developer of liveness facial biometrics in the country, with the rest of the market reselling technology built elsewhere. That has consequences for accuracy, because darker skin tones reflect less light than lighter ones, and a model trained elsewhere carries biases based on the data it is trained on.
“Our datasets are more honed into that because people are using our facial biometrics,” he adds. Building locally also means a workflow can be reshaped around a client instead of sold as it comes, and the development behind it is charged in rands rather than dollars.
“That's where the world is going, and we're making it possible for a normal provider selling a service to do it in seconds,” he concludes.
You can view the full video here.
* ITWeb TV Biz episodes are sponsored.
Share
