Face ID Under the Skin: The 30,000-Dot Anatomy of a Face Scan

01 Why Passwords and Fingerprints Were Not Enough

For most of the history of personal computing, the only thing standing between your data and an intruder was a string of characters. Passwords scale badly for a device you unlock dozens of times a day: people reuse them, write them down, and hand them to phishing pages. PINs on phones shortened the friction but kept the weakness - four digits offer only 10,000 combinations, and the smudges on the glass sometimes betray the pattern itself.

Fingerprints looked like the answer when Apple shipped Touch ID in 2013. A print is convenient, unique, and always with you. But it has structural problems. You leave copies of your fingerprint on everything you touch - glasses, door handles, the phone screen itself - and once a print is compromised you cannot reissue it the way you rotate a password. Wet or greasy fingers misread, and gloved hands cannot use it at all.

Apple's published security statistics frame the upgrade Face ID was designed to deliver. Touch ID's false-accept rate was quoted at 1 in 50,000: the chance that the fingerprint of a random person could unlock your phone. Face ID's figure is 1 in 1,000,000, roughly twenty times stricter, achieved with no more effort than looking at the screen. Those are population-level odds, not a guarantee against a determined attacker who specifically targets you - a distinction this article returns to below.

The numbers also explain why the industry kept pushing beyond prints. If the false-accept gap is real, the face is simply a longer key - and unlike a password, it cannot be forgotten or left at home.

02 The TrueDepth Stack: Projector, Flood, and Camera

The hardware that makes this possible hides in the notch, and it is far more than a camera. The TrueDepth module packs a flood illuminator, a dot projector, an infrared camera, a proximity sensor, an ambient light sensor, and a conventional front camera into a strip a few millimeters tall. The three components that do the security work are the illuminator, the projector, and the IR camera.

The flood illuminator bathes the face in infrared light the eye cannot see. That solves darkness first: Face ID works in a pitch-black bedroom precisely because it carries its own light. The dot projector then fires out more than 30,000 invisible infrared dots - Apple's published figure - arranged in a known pattern. The infrared camera watches how that pattern lands. On a flat surface the dots sit undisturbed; on a nose, a cheekbone, or a glasses frame they stretch, compress, and shift in ways that encode the three-dimensional shape beneath them.

This is the technique called structured light: project a known pattern, observe its deformation, and compute geometry from the distortion. It is the same principle industrial 3D scanners use, miniaturized to fit a phone. To an ordinary camera the whole exchange is invisible, which is why the source video - shot through an infrared camera - is such a useful demonstration: it shows the unseen dot grid as a constellation moving across the face with every attempt.

03 From Dot Grid to Mathematical Face

Each of the 30,000 dots lands somewhere the projector did not quite predict. The software compares the expected pattern with the observed one, and the displacement of every dot becomes a depth reading. Stitched together, those readings form a depth map - a grid of distance measurements, one for each sampled point on the face. Where a photograph records brightness, the depth map records geometry.

From that depth map the system computes a mathematical representation of the face: a set of numbers describing its three-dimensional structure, not a stored picture. Authentication is a comparison between two sets of numbers - the live capture and the enrolled model. This is why Apple can state that the enrolled data is not a photo of your face; by the time it exists, the image has already been reduced to geometry.

Enrollment is quick because the model is built to keep learning. Every successful unlock feeds the enrolled model, so it gradually tracks a growing beard, new glasses, seasonal weight changes, and aging. The enrollment is not a frozen snapshot from the day you set up the phone - it is a moving average of who you currently are.

04 The Neural Engine and the Secure Enclave

Turning a distorted dot pattern into a compact numerical model of a face is exactly the kind of task neural networks were built for. Since the A11 Bionic chip, iPhones have carried dedicated silicon for this: the neural engine performs the conversion and the comparison against the enrolled model on hardware purpose-built for the job, in a fraction of a second and at very low power.

The security architecture around that computation matters as much as the computation itself. The enrolled model lives inside the Secure Enclave, a coprocessor with its own encrypted memory that the main operating system cannot read. Apps that request Face ID authentication never receive the face model, or even the result of the comparison - the system hands them a yes or no. If the phone is compromised, the face data stays sealed in a part of the silicon the attacker cannot address.

On top of matching, the system runs attention checks: it verifies that your eyes are open and directed at the sensor. This is a liveness signal as much as a usability feature - a photograph held in front of the camera should fail not only because it is flat, but because it is not looking back at the phone. The requirement can be relaxed for accessibility, and each relaxation is documented to trade security for convenience.

05 Why Face ID Fails: Angles, Hats, and Sunlight

Most Face ID failures are false rejections, not false accepts, and the causes are mundane. Holding the phone too far below your face - on a desk, in a car mount - pushes the capture outside the angles the system was trained on. A wide-brimmed hat or a scarf changes the visible geometry. Strong direct sunlight floods the sensor with infrared and can drown out the projected dots, a limit of any active illumination system.

The pandemic ran this experiment at planetary scale: face masks broke Face ID everywhere at once. Apple's eventual answer, mask unlock in iOS 15.4, authenticates on the geometry of the eyes alone. It works, but it is explicitly weaker - Apple documented the reduced security - and it is best understood as the pipeline running on a fraction of the data it was designed for.

Wet screens are a different story. Condensation on the glass can scatter the projected dots, but most frustration near water is the touchscreen refusing a wet swipe, not the face scan failing. The distinction is worth making: a scan that fails is an annoyance; a scan that accepts the wrong person would be a vulnerability. Documented real-world failures are overwhelmingly of the first kind.

06 Spoofing, Liveness, and the Privacy Debate

Attacking Face ID head-on has a history. Photographs fail immediately - they have no depth. Simple masks fail because their surfaces distort the dot grid in ways a real face does not. The best-publicized break, by the Vietnamese security firm Bkav in 2017, used a purpose-built 3D mask combining sculpted features with infrared-printed regions and reportedly defeated the first-generation system. The lesson is about threat models: the 1-in-1,000,000 figure describes random strangers, not an adversary who spends real money on a mask of your specific face.

The liveness counterweight is attention detection. A statue does not look at the phone; demanding an open-eyed gaze forces spoofing attempts to simulate behavior, not just shape. That said, the system has no heartbeat or thermal requirement, and researchers continue to probe the boundary. Apple's figures should be read as population statistics with a strong record against casual spoofing - not as an absolute guarantee.

The privacy debate runs on two tracks. The first is data handling: Apple states that all face data is processed on the device and never leaves it, and the Secure Enclave architecture makes that a verifiable design claim rather than a promise. The second is the nature of biometrics themselves. A password can be changed after a breach; a face cannot. And in some legal systems, courts have treated compelled biometric unlocks differently from compelled passcodes - police can sometimes order a suspect to hold a phone up to their face where forcing a spoken password would not be allowed. Face scanning is thus both the most private authentication most people use daily and, legally, one of the least defensible.

07 What Comes After the Notch

The TrueDepth stack is a victim of its own success: the notch that houses it became the phone's most criticized feature. The industry's direction is under-display sensing - moving the projector and the infrared camera behind the screen itself, the way under-display fingerprint readers have already shipped. The engineering problem is hard, because the display layer in front of the dot projector must let a clean infrared pattern through without scattering it.

Competing depth-sensing designs are evolving too. Time-of-flight sensors, common on some Android flagships, replace the projected dot grid with a light pulse and measure the round trip of its reflections - coarser coverage, simpler optics. Cheaper phones use purely two-dimensional face recognition from an ordinary camera, which a printed photograph can sometimes defeat. Structured light remains the most secure approach in the field, and its false-accept record is the benchmark the others are chasing.

The plausible endpoint is a phone with no visible sensor cutouts at all, authenticating with the same million-to-one odds through a screen that betrays nothing. The dots would then be invisible in a second way - woven into the display, one more layer of the scan hiding under the skin of the device itself.