A little more than six years ago, Apple unveiled Face ID. It was a new method to biometrically unlock iPhones and authenticate purchases by scanning your face. Yet after all this time, there still hasn’t been a meaningful competitor on Android—at least, not with the same level of security and capabilities.
Google’s Pixel 8 has Face Unlock, but it has trouble working in the dark; the Face Unlock available on Samsung smartphones can’t be used for secure applications, such as banking. In Androidland, the fingerprint scanner is king, but that might not be the case for long.
Metalenz, a startup pioneering optics technology called “optical metasurfaces,” is hard at work on introducing secure face authentication to Android with its Polar ID technology. Late last year, it announced a partnership with Qualcomm to port its ongoing development to the chipmaker’s flagship processor. Today, at Mobile World Congress 2024 in Barcelona, it announced that it will be using Samsung’s Isocell Vizion 931 image sensor to power its imaging system.
I visited Metalenz’s headquarters in Boston to get a first look at Polar ID. The system is still in its early stages, and the company is currently gathering large amounts of data to improve its facial recognition machine learning algorithms. But it has plans to send development kits to smartphone manufacturers in the middle of this year for testing, which means there’s a good chance we’ll see a Face ID–like system for Android, one that’s potentially better than Apple’s approach, inside smartphones by early 2026.
New Optics
Metalenz is a startup born out of a research group at Harvard University and was founded by CEO Robert Devlin and physicist Federico Capasso. I’ve been covering its development ever since the company emerged from stealth mode in 2021. That’s when it unveiled its metasurfaces technology—a flat-lens system that takes up far less space than the traditional multi-lens elements used in most smartphones today.
The iPhone 14 Pro, for example, has seven lens elements (layers of glass or plexiglass) stacked above the camera sensor. Having multiple lens elements improves image clarity, captures more light, and corrects issues such as chromatic aberration (where colors are visible on the fringes of images). But it also adds more complexity and requires more space in the smartphone. Metalenz’s metasurfaces is a single lens that uses nanostructures to bend light rays toward the camera sensor, performing the job of multiple lens elements in a much smaller package.
A wafer of multiple metasurfaces, the flat-lens system takes up far less space than multi-lens elements used in most phones currently.
Photograph: Julian Chokkattu
In 2022, the company showed off the capability of its metasurfaces to capture polarization data. When light hits an object with shape and specific material properties, it will have polarization—every object creates a unique polarization signature. To quote my own words again: “The waves of light that bounce off of regular ice on the side of the road will oscillate differently than the light that bounces off black ice. If a camera can pick up this information, you can feed it to a computer-vision machine-learning algorithm and train it to learn the difference between black ice and normal ice. Now the car can advise you of the oncoming danger.”
That brings us to Polar ID. Metalenz has been working on a system that can authenticate a person’s face based on their polarization signature. By tying this signature with facial recognition, the system can recognize and authenticate an individual whether they are wearing a surgical face mask or sunglasses, or even when someone is trying to spoof your face with a picture or 3D mask. It just needs to match the polarization signature of your skin.
Apple’s approach in Face ID uses a structured light transmitter, a pricey module that also takes up a lot of space (hence the chunky notch, or pill, at the top of the iPhone’s display). This technology is derived from PrimeSense, a company Apple acquired in 2013 (the very same that was behind the 3D sensor in Microsoft’s Kinect). Devlin says Apple holds several key patents around the technology, which is one of the reasons there is a high barrier for competitors to wade in. It’s also just too complicated and expensive for Android manufacturers to replicate the experience, though a few have tried over the years.
“The Android community has still really wanted this,” Devlin says. “They want it so bad that they’re now putting out substandard solutions so people can unlock, and oftentimes the user doesn’t know that the security isn’t there.”
Apple’s Face ID system first takes a 2D image of your face with the TrueDepth selfie camera—this is to verify the person in the image is you, based on the images you provided Apple when setting up Face ID. Then the structured light transmitter shoots out 30,000 infrared dots on your face to create a 3D rendering and match your face’s contour. If your face is obscured, like if you’re wearing a surgical face mask, there are fewer data points, reducing the accuracy.
The Polar ID metasurfaces combined with Samsung’s Isocell Vizion 931 sensor.
Photograph: Julian Chokkattu
With Polar ID, Devlin says the system has 10 times the resolution—400,000 pixels. It’s worth noting that Polar ID doesn’t replace a phone’s selfie camera but instead sits next to it. Traditional cameras can’t capture polarization data, but the metasurfaces can pull in this information and send it to the image sensor to reconstruct it. There’s no need for a structured light module. Instead, Polar ID uses a flood illuminator, which uses near-infrared light to capture a picture of your face in any lighting condition. The system verifies that the 2D image looks like you, and then it analyzes the same image’s polarization signature to make sure it matches what it has stored.
This is one less step than Apple’s approach, and at significantly reduced price (and space). Devlin says the structured light transmitter Apple uses is around $14, a high cost on its bill of materials (the cost of individual components that make up the phone). It might not sound like much, but Metalenz’s system sits at roughly around $5 right now in a small volume. If Apple used it instead, that’s potentially billions of dollars in savings.
Metalenz’s system can also be put together in the same chip fabs where a phone’s processors and image sensors are manufactured, further streamlining the supply chain.
Laser-dot pattern projectors are one of the most complex and expensive modules in phones. Polar ID’s tech can supposedly recognize and authenticate a person with a single image.
Courtesy of Metalenz
Metalenz’s optics are already in today’s smartphones. It partnered with STMicroelectronics in 2022 to integrate metasurfaces with the company’s time-of-flight sensors—these are common sensors in smartphones used to capture depth information to improve features like portrait mode.
Teardowns have confirmed phones like the Google Pixel 8 and Samsung Galaxy S23 are using Metalenz’s optics. What used to require eight lens elements before now uses just two—one to send out the laser to measure distance, and one to receive it. Metalenz claims its product improves the range of depth sensing and its accuracy. The company is also working with Dilusense in China for its payment kiosks, which primarily use facial recognition for payments.
Face Off
Polar ID authenticates a face based on its polarization signature. It supposedly works when an individual is wearing a surgical face mask and sunglasses, and stops ID attempts with a picture or 3D mask.
Photograph: Julian Chokkattu
At Metalenz’s office in Boston, I sat in a room filled with several 3D masks of people’s faces, something you’d expect to find in Mission Impossible. The company had its Polar ID system connected to a smartphone, complete with the Samsung Isocell Vizion 931 sensor, which Devlin says is optimized for the wavelengths needed to capture polarization data.
You register your face much like how you’d set up Face ID on an iPhone. Rotate your face in a circular motion, allowing the system to capture a variety of angles. That’s it. If manufacturers partner with Metalenz to license its technology, the setup process will likely be customized to match the brand’s software (and probably won’t be called Polar ID).
Devlin went through a few accessories to show the system’s efficacy. He donned a high-quality face mask of himself and Polar ID recognized it as a spoof. Devlin says in the company’s tests, it was able to access a locked Pixel 8 using the same high-quality 3D masks made from 3D models of the face. Metalenz says this “falls in line” with Google’s security specs, whereas Apple’s security standard requires that 3D scans are rejected. But when Devlin put on a surgical face mask, Polar ID didn’t have any trouble recognizing him.
It was my turn. I scanned my face and … Polar ID had trouble recognizing me. Turns out the company placed some limitations on the system—it was able to recognize my face in a limited field of view, but it struggled to work when I donned a 3D mask of my face, and when I wore a surgical mask.
Devlin says the current model has parameters to “err on the side of caution,” supposedly explaining why it didn’t work too well for me. The company at present also doesn’t capture as many facial angles as Face ID in its enrollment process, though that will change as development continues.
This is coupled with the fact that the company lacks a diverse data set of faces to train its algorithms on. “For a very robust facial recognition model, the data set and diversity of images we have worked with to date is on the smaller side compared to what existing RGB face recognition models have,” Devlin says. Metalenz is working with third-party vendors and OEMs to collect more so that its algorithms improve.
Photograph: Julian Chokkattu
In a way, the lack of data—especially polarization data—is partly why Metalenz focused on facial recognition first. By deploying its Polar ID imaging system in smartphones, it can tap into a broad dataset, giving it a stronger foothold to wade into other categories with its polarization technology.
You could, theoretically, use the camera on your phone to check your skin for cancerous growths or look at the particulates in the air around you to measure local air quality.
“Even glucose has a very specific polarization signature,” Devlin says. “We see a whole bunch of other applications. That’s the nice part about facial recognition being the beachhead … It’s the largest data set of polarization information of people that has ever been out there. The applications that will emerge from that is another thing that we think is really interesting about putting polarization in devices for the first time.”
Devlin reiterates that all of the data Polar ID captures is saved in the trusted execution environment on Qualcomm’s chipset, where it’s encrypted. If the phone manufacturer and user agree to share data, then every face unlock can be used to improve and expand Metalenz’s database. The system will also improve over time for the user on the device the more it’s used.
But, hypothetically, what if your unique polarization signature gets out there? If a public security camera is equipped with the same polarization-gathering technology, it’s possible that it can identify you. Devlin says the camera would need to have a high-enough resolution at those distances to really work, and believes it’s unlikely.
If and when Android phones start to adopt this technology, Devlin believes it will spell the end of the fingerprint sensor, which he says is still not as secure as facial recognition. “As we get this out there and continue to refine it, we can have models that could even be competitive in price with low-end capacitive sensors.”
