Key takeaways
- Face recognition verifies or identifies people by comparing deep learning face embeddings.
- It began in the 1960s and leapt forward with deep learning from 2014.
- Faces are biometrics, while passkeys are cryptographic credentials that face unlock merely authorises.
- YOLO detects faces, but recognition needs embedding models such as ArcFace.
- Spoofing and demographic bias are real risks, so use liveness checks, testing and consent.
Face recognition identifies or verifies people from images of their faces, using deep learning models that turn each face into a numeric embedding. It powers phone unlocking, photo organisation and security systems, and raises real questions about spoofing, bias and privacy. I built FaceVision, which runs detection, recognition and liveness in the browser and stores only embeddings, so this page covers what face recognition means, how the AI works, its history, biometrics, passkeys, YOLO, spoofing, bias and Google.
What does face recognition mean?
Face recognition is technology that identifies or verifies a person by analysing their facial features in an image or video. Verification checks whether a face matches one claimed identity; identification searches a database for who the person is.
It differs from face detection, which only finds where faces are, and from facial analysis, which estimates attributes such as expression. Verification, one-to-one matching, powers phone unlocking and identity checks. Identification, one-to-many matching, powers photo tagging and surveillance, which carries greater privacy risk. Most systems follow the same pipeline: detect the face, align it, convert it into an embedding and compare embeddings. FaceVision follows this pipeline entirely in the browser.
How does face recognition AI work?
A detector finds and aligns the face, then a deep neural network converts it into an embedding, a vector of numbers capturing identity. The system compares embeddings using cosine similarity and declares a match above a threshold.
Networks are trained on millions of labelled face images with loss functions such as ArcFace, which push embeddings of the same person together and different people apart. The threshold controls the trade-off between false matches and false rejections, chosen for each application's risk. Liveness detection checks that a real person is present. FaceVision runs these models with ONNX Runtime Web in the browser and stores only embeddings in PostgreSQL, never the original images, which limits privacy exposure.
Does face recognition use AI?
Yes. Modern face recognition relies on AI, specifically deep learning with convolutional neural networks or vision transformers. Earlier systems used hand-designed features and statistical methods, which were far less accurate.
The shift happened around 2014, when deep learning systems such as Facebook's DeepFace and Google's FaceNet reached near-human accuracy on benchmark datasets of faces photographed in uncontrolled conditions. Today, AI handles every stage: detection, landmark alignment, embedding, liveness checks and quality assessment. Phone face unlock combines AI with specialised hardware, such as infrared and depth sensors on iPhones. The AI runs on the device in phones, keeping face data local rather than sending it to servers.
When did face recognition start?
Face recognition research began in the 1960s, when Woody Bledsoe and colleagues manually marked facial features for computers to compare. Automated systems followed in the 1970s, and the Eigenfaces method in 1991 marked a major step.
In the 1990s, the United States government's FERET programme built datasets and benchmarks that accelerated research. The Viola-Jones detector in 2001 made real-time face detection practical, appearing in cameras and phones. Deep learning transformed accuracy from 2014 onward, with DeepFace, FaceNet and later ArcFace. Face unlock reached mass-market phones with Apple's Face ID in 2017. Since then, regulation and debate over surveillance and bias have grown alongside the technology's spread.
Is face recognition a biometric?
Yes. Face recognition is a physiological biometric, identifying people by physical characteristics, like fingerprint, iris and voice recognition. Many privacy laws treat biometric data as sensitive, requiring consent and strict protection.
As a biometric, it is convenient because it needs no contact or remembered secret, but faces are public and cannot be changed if data leaks, unlike passwords. That makes secure storage critical: phones keep face templates in secure hardware, and well-designed systems store embeddings rather than raw images, as FaceVision does. Regulations such as the EU's GDPR classify biometric data used for identification as special category data, and India's data protection law also governs its processing.
Is face recognition a passkey?
No. A passkey is a cryptographic credential stored on your device, based on the FIDO2 and WebAuthn standards, that replaces passwords. Face recognition is often how you unlock and approve using that passkey on your phone or laptop.
When you sign in with a passkey, the website sends a challenge, your device signs it with a private key that never leaves the device, and the site verifies it with your public key. Face ID, fingerprint or a device PIN simply confirms that you are present to authorise the signature. Your face data never goes to the website. This design resists phishing far better than passwords, since passkeys only work on the genuine site.
Can YOLO do face recognition?
YOLO can detect faces, finding where they are in an image, when trained or fine-tuned on face data, but it does not recognise who someone is. Recognition requires an embedding model, such as ArcFace, applied to each detected face.
A common pipeline uses a YOLO face detector to locate faces quickly in video frames, crops and aligns each face, then passes it to a recognition model that produces embeddings for comparison with a database. YOLO's speed suits real-time detection on many faces, while dedicated detectors such as RetinaFace or YuNet provide landmarks useful for alignment. Respect consent and local laws, especially for recognition in public or workplace settings.
Can face recognition be fooled?
Yes, depending on the system. Simple 2D systems can be fooled by printed photos, screen replays or masks, while 3D systems with liveness detection are much harder to trick. Identical twins and deepfakes pose further challenges.
Researchers have also shown adversarial patterns, such as specially designed glasses or makeup, that confuse some algorithms. Defences include depth and infrared sensing, liveness checks such as blink or motion prompts, texture analysis and challenge-response tests. FaceVision includes liveness checks for this reason. For high-security applications, combine face recognition with another factor, and regularly test systems against new spoofing techniques, since attackers adapt quickly as generative AI improves.
Does face recognition have bias?
It can. NIST's 2019 study of nearly 200 algorithms found that many had higher false match rates for some demographic groups, varying by algorithm, while the most accurate algorithms showed much smaller differences.
Earlier, the Gender Shades study in 2018 found commercial gender classification systems performed worst on darker-skinned women, prompting vendors to improve. Bias usually comes from unbalanced training data and evaluation sets. Mitigation includes diverse datasets, testing error rates per group, choosing thresholds carefully and keeping humans in the loop for consequential decisions. Because errors can lead to wrongful accusations, many jurisdictions restrict police and public-space uses of face recognition.
Can Google do face recognition?
Yes, in specific products. Google Photos groups similar faces in your library where Face Groups is available, and Pixel phones offer face unlock. Google Cloud Vision detects faces and expressions but deliberately does not identify who people are.
Google Lens also avoids identifying people from photos, reflecting Google's policies on facial recognition. For developers needing recognition, options include open-source models, such as ArcFace implementations, and cloud services from other providers, used with consent and legal review. Google's ML Kit offers on-device face detection with landmarks and contours for apps, a useful first stage before recognition. Always check regional availability, since face features vary by country because of privacy laws.
Sources
- NIST, Face Recognition Vendor Test Part 3: Demographic Effectsnvlpubs.nist.gov
- Buolamwini and Gebru, Gender Shadesproceedings.mlr.press
- Deng et al., ArcFace: Additive Angular Margin Loss for Deep Face Recognitionarxiv.org
- FIDO Alliance, Passkeysfidoalliance.org
- Google Cloud, Detect faces with Cloud Visioncloud.google.com
- Wikipedia, Facial recognition systemen.wikipedia.org
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
