A recent smartphone teardown highlighted a familiar engineering move: shrinking the visible sensor cutout by placing part of the face authentication system under the display. The deeper concept is not the cutout itself, but how Face ID turns optics, machine learning, and secure hardware into a practical identity check.
Why this matters now
Face authentication has become a default interaction pattern for phones, payments, password managers, workplace devices, and emerging wearable interfaces. Professionals should understand it as more than a convenience feature. It is a compact identity system with security, privacy, usability, and hardware tradeoffs.
The under display trend makes those tradeoffs more visible. Product teams want uninterrupted screens, but sensors still need light, space, calibration, and processing. Moving an infrared camera behind a display means the panel must show normal pixels to the user while also allowing enough infrared signal to reach the sensor. That is a systems problem, not just industrial design.
For AI and tech learners, Face ID is a useful case study in applied machine learning: sensing, feature extraction, matching, adversarial resistance, latency constraints, and on device privacy all have to work together.
How it works
Face ID is a biometric authentication method that verifies a person by comparing a live facial measurement against an enrolled template stored securely on the device. The important point is that the system is not simply taking a selfie and comparing photos. It uses specialized sensing, often infrared light and depth information, to build a more robust representation of face geometry.
Face ID authentication flow
Enrollment ·························
│
▼
Infrared sensing ···················
│
▼
Depth map and features ·············
│
▼
Secure match ·······················
│
▼
Unlock or reject ···················
A live infrared measurement is matched against a secure enrolled template.
During enrollment, the device captures multiple views of the user’s face and converts them into a mathematical template. That template is typically stored in protected hardware and is not meant to leave the device.
During authentication, the system illuminates the face with infrared light, captures the reflected pattern, estimates depth, and extracts features that are stable enough to recognize the user but flexible enough to handle normal changes such as facial hair, glasses, or lighting. A machine learning model helps compare the live features with the enrolled template. If the match score crosses a threshold, the device unlocks or authorizes an action.
Under display Face ID adds another layer of complexity. The display region above a sensor must balance two jobs: emitting visible light for the screen and transmitting infrared light for sensing. A special pixel pattern can create a path for infrared while trying to remain visually normal. The tradeoff is engineering complexity: lower signal, more calibration, potential display artifacts, and tighter manufacturing tolerances.
Real-world applications
The obvious application is device unlock, but the same pattern extends to payment approval, password autofill, app access, workplace sign in, and high assurance consumer identity flows. In each case, face authentication reduces friction by replacing something the user knows, such as a password, with something the user is.
It also appears in adjacent domains: access control for secure rooms, driver monitoring, augmented reality headsets, and healthcare workflows where hands free authentication is valuable.
The risks are equally important. Systems must defend against photos, masks, replay attacks, and coerced unlock attempts. They must handle false rejects without locking out legitimate users and false accepts without weakening trust. Good implementations combine biometric matching with secure hardware, rate limits, liveness checks, and fallback credentials.
Where to go deeper
To build a stronger mental model, study on device AI, secure hardware, computer vision, and mobile system design. EducationPals learners can connect this topic to Arm big.LITTLE for understanding mobile performance and power tradeoffs, and to Android sideloading for the broader security model around trusted and untrusted apps.
If you are moving toward AI product or platform work, also study text embeddings, vector databases, and retrieval-augmented generation. They use different data types, but the transferable idea is similar: convert messy real world inputs into mathematical representations, compare them efficiently, and design the surrounding system so the result is useful, secure, and reliable.
