Diversity of data: The key to secure 3D face authentication

Big is not always best

One of the major factors that too often gets overlooked when tech companies are developing biometric-based technologies — especially ones that use face recognition for authentication — is the fact that we are all now operating in a global business community. A quite telling article in the New York Times not long ago by Steve Lohr basically decried the current state of mobile face recognition. The reporter’s key takeaway on the present situation was that it works well if you are a middle-aged white male.

Given that the world is flat and continues to get flatter (great book by Thomas Friedman, btw), this is not an acceptable situation. People on the planet have faces of all different shapes, sizes and colors. Plus, tech research firm Gartner predicts that by 2021, 40% of smartphones will be equipped with 3D cameras. Add to this the fact that the number of mobile phone users is forecast to reach 4.77 billion in 2017 and pass the five billion mark by 2019. The net impact of inconsistent, face-based authentication could be significant, as people increasingly expect to easily access their devices using this approach.

Lots of diverse data to the rescue

Having worked in this space since 2006, I have to say that the inconsistent performance of both standard 2D and the new 3D face authentication is too often a result of simply defined or poorly curated databases. This is one of the reasons why face authentication has struggled to achieve mainstream acceptance.

Delivering accurate, consistent and fast 3D face recognition requires the factors I mentioned above: exploiting today’s tiny and accurate cameras and using advanced AI algorithms to capture, manage and rationalize the torrent of data generated. For perspective, today’s state of the art 3D cameras capture 30K-40K data points per scan every time they look at a face — but that is not all.

Image source: WLFDB

Many companies claim that because they have a large database, their technology is going to be accurate. Which is, frankly, inaccurate. More important than size is the diversity of the database for the specific intended use cases. Millions don’t matter. Databases can in fact be statistically significant with as few as 250 persons if the appropriate factors are captured.

In order to deliver an acceptable level of 3D accuracy, the data needs to:

• Include people of different races and different genders, with and without glasses;
• Represent people in different positions;
• Be acquired over long periods of time;
• Be captured with the same types of cameras used for recognition on various end user devices; and
• Represent expected use case environments — for example, will people be only sitting in a well-lit office or might they be lounging on the beach in bright sunlight?

My recommendation has always been to focus on quality and diversity of data rather than just the size of the database. Over the past 12 years, our team has been able to generate and use a database of millions of people. More importantly, it includes images of users of different races from all over the world acquired with cameras and in environmental conditions and poses representing how they are actually using the technology in real-world situations.

I encourage organizations interested in exploiting the power of 2D or 3D face recognition and authentication to be sure their proposed approach has access to a meaningful and diverse data set. It will ensure consistent results and, in turn, help drive overall customer satisfaction when deploying a face authentication system.

All IoT Agenda network contributors are responsible for the content and accuracy of their posts. Opinions are of the writers and do not necessarily convey the thoughts of IoT Agenda.