acm-header
Sign In

Communications of the ACM

ACM News

Democratizing Facial Authentication


View as: Print Mobile App Share:
Intels RealSense facial authentication module, shown for comparison purposes next to a credit card.

Intels RealSense facial authentication module (shown here next to a credit card) is small enough for almost any Internet of Thing device, as well as for a wide variety of other applications.

Credit: Intel

Intel demonstrated facial authentication solutions at the most recent Consumer Electronics Show in January (CES 2021), which the company said are as accurate as Apple's Face ID for iPhone, claiming one false positive in a million identifications. Intel's similarly named RealSense ID technology performs face authentication for any device running the Windows, Linux, or Android operating systems.

Jim McGregor, principal analyst at Tirias Research, a market research firm focusing on "technologies, markets, and ecosystems," said the technology inside Intel's solution is more versatile, and its target applications are much wider, than those of Apple Face ID. "Android smartphones will definitely get face authentication this year, and Intel is not the only source. Also, for Intel there are many other edge-device applications for RealSense ID that use Windows, Android, and Linux operating systems," said McGregor. "In fact, by 2025, every network edge device will be using AI (artificial intelligence) components."

Intel's RealSense ID contains its own neural network to accelerate its smart AI functions "on device"—that is, without requiring a network connection. Intel is not yet selling its chip set, just a full-fledged USB peripheral in a 62-by-32.5-by-11 millimeter case (model F455), and for integration inside vendor cases, a printed circuit board (PCB) module measuring 50-by-18-by-4.6 millimeters (model F450).

How It Works

Like Apple's Face ID, Intel's RealSense ID projects invisible infrared dots onto the face of a user (Apple uses up to 30,000 dots, while Intel does not reveal how many dots it is using). The similarities pretty much end there. Intel's RealSense ID device uses two cameras to create a depth map of the face by measuring the stereo parallax between those dots. Apple's TrueDepth camera, on the other hand, measures the time of flight between the iPhone's emission of infrared dots projected onto the user's face and their reflection back to the TrueDepth camera.

According to McGregor, Intel's RealSense ID's two-camera approach creates a depth map that is not only more accurate than time-of-flight calculations, but is potentially much more versatile than Apple's Face ID. "Apple's Face ID is more cost-effective, but authentication is all it does," said McGregor. "Intel's use of two cameras, on the other hand, provides not only more accurate 3D mapping, but also provides applications with all sorts of metrics, such measuring the sizes of rooms, measuring the sizes and shapes of objects, making super-accurate maps of almost any kind of space, and many automotive applications, such as alerting inattentive or drowsy drivers. RealSense technology can do this because it is based on a decade of development at Intel where they already sell a line of stereo camera-based devices."

Intel also claims RealSense ID has superior "anti-spoofing" capabilities to Apple's Face ID. Intel's RealSense ID combines its cameras with an infra-red (IR) illuminator, to keep from being fooled by two-dimensional (2D) photographs, videos, or even sophisticated three-dimensional (3D) face masks.

Joel Hagberg, head of marketing and product management for the Intel RealSense line, said Intel RealSense ID "is designed to generate face templates [composed of landmarks rather than images] that are encrypted and stored securely on-device." RealSense ID also bases its security algorithms on a hardware root of trust that cannot be hacked, as software can. The entire device is also based on a purpose-made system-on-chip (SoC) proprietary to Intel.

RealSense ID, says Hagberg, also claims advantages over other facial authentication systems that include features to prevent their use by governments for surveillance applications. In particular, Intel's RealSense ID will not run its authentication algorithms unless physically prompted by a pre-registered user of a particular device.

The technology adapts to users over time as their physical features change, including an individual's hair and glasses. The system uses "a custom neural network that has been trained on a large proprietary data set that spans different skin colors, age groups, nationalities and gender," said Hagberg. "Intel is committed to respecting human rights and avoiding complicity in human rights abuses. Intel's products and software are intended only to be used in applications that do not cause or contribute to a violation of internationally recognized human rights."

Another unique capability of RealSense, according to Hagberg, is an algorithm that allows original equipment manufacturers (OEMs) to adjust the confidence levels of its technology in order to optimize it for specific use-cases, such as securely authenticating similar-looking family members, siblings, or even identical twins. (Apple, on the other hand, warns Face ID's "statistical probability is different for twins and siblings.")

Smartphone Authentication

In preparation for integrating facial authentication solutions into Android smartphones, Google has prepared Android versions 10 and above to support a special hardware abstraction layer (HAL) of software that provides access to hardware facial authentication resources through programming interfaces.

Apple, Intel, and other manufacturers plan to increase cybersecurity for all edge devices by adding multiple biometric factors, said McGregor, including retina scans and earlobe characterization. Retina scans already are being used for "top secret" access to secure areas, but it turns out the size, shape, and contours of each person's earlobe are as unique as their fingerprints and retina patterns. What's even better is that you don't need a new kind of sensor to scan earlobes—a 3D camera works just fine.

"We're still in early days of AI, which is essential for using any biometric sensors. Neural networks and deep learning algorithms are all improving rapidly to meet this challenge; for instance, you can add a heartbeat sensor to a fingerprint sensor to make sure the finger is not severed," said McGregor.

"What Intel is doing with RealSense, so far, is an adaptation of current technologies. However, Intel's 12th generation processor, code-named Alder Lake, will raise the bar with heterogeneous AI and neural network cores right on its CPU chip."

R. Colin Johnson is a Kyoto Prize Fellow who ​​has worked as a technology journalist ​for two decades.


 

No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account