Technology review
A comparative review of eleven methods, from a paper signature to face recognition — based on documented cases and independent measurements, not on marketing claims.
Taking attendance looks, from a distance, like a solved problem. A list, a signature, done. In reality every attendance system has to answer the same hard question, and most of them cannot: is this person really here, and is it really this person?
These two questions are independent of each other, and most technologies solve only one. When a QR code is scanned, the system knows that the code was scanned — it does not know who scanned it. When a fingerprint is read, the system knows who read it — it does not know whether that person stayed for the rest of the class. The history of attendance systems is largely the history of trying to close these two gaps.
In this article we examine eleven methods one by one. For each we ask three things: how well does it work, how does it break, and is there documented evidence that it has broken. In the last section we describe where yoklama.ai sits in this table — with its strengths and its weaknesses.
We use six criteria to compare the methods. They are independent of each other; a method can be excellent on one and a disaster on another.
| Criterion | Question | Why it matters |
|---|---|---|
| Identity assurance | Is the person checking in really that person? | The only antidote to proxy attendance |
| Presence assurance | Is the person really in the room? | Prevents checking in remotely |
| Continuity | Did they stay for the whole class? | Prevents the “check in and leave” trick |
| Privacy | How much sensitive data is processed? | Compliance with Turkish data protection law (KVKK) and the GDPR; institutional risk |
| Friction | How much does it disrupt the flow of the class? | The decisive factor in adoption |
| Accessibility and cost | Can everyone use it, and what does it cost? | Students with no phone, disabled students, hardware budget |
A sheet is passed around and everyone signs next to their name. Or the teacher reads out the names and the students answer.
Assessment: Unmatched on privacy, the weakest method on security. Signing on a friend's behalf is so entrenched in student culture that it is the first thing that comes to mind when anyone says “attendance fraud”. Even so, it has to exist as every system's fallback: the student with no phone, a broken system, a power cut.
The student presents a card to a reader. Depending on the card type, a barcode scanner, a magnetic reader or an RFID/NFC antenna is used. In Turkey, university campus cards are mostly from the MIFARE family.
Assessment: MIFARE Classic's Crypto-1 algorithm was broken in ESORICS 2008 and IEEE S&P 2009; in 2024 a hardware-level backdoor was found in some Chinese-made clones (IACR ePrint). Cloning a card with a $20 Proxmark device is a routine operation today. In practice the identity assurance of card-based attendance is only slightly better than a paper signature — and a cloned card leaves less of a trace than a forged signature.
The teacher shows a QR code on screen or on the projector; students scan it with their phones. There are two sub-types: static QR (the same for the whole class) and dynamic/animated QR (renewed every few seconds).
Documented case: In October 2025, in the Data Science Discovery course at the University of Illinois Urbana-Champaign, attendance was taken with a QR code plus a 90-second question. The students in the room told their friends outside that the question had been posted; those who were absent answered remotely. In a class of about 1,200, more than 100 students were identified — the instructors caught them by examining page-refresh patterns, IP addresses and server logs. This case shows two things at once: static/single-window QR can be broken, and log-based forensic analysis really does work.
The teacher writes a short code on the board or on screen; the students type it into the app.
Assessment: Indiana University's official guidance for Top Hat, after acknowledging confirmed cases of academic misconduct, makes a concrete recommendation: show the code for no more than 10 seconds. When a method's own official usage guide has to say “show it as little as possible”, that says enough about the method.
Each student has their own small radio-frequency response device. A question is asked in class and answered with the device; participation counts as attendance.
Documented case: In January 2015 at Dartmouth College, 64 students were charged with honour code violations in the “Sports, Ethics and Religion” course — they had been bringing absent friends' clickers to class and answering on their behalf. The University of Arkansas's academic integrity pages also describe the same act explicitly as cheating and document cases of students carrying more than one clicker.
The teacher's device, or a beacon placed in the room, broadcasts a Bluetooth Low Energy signal. The student's phone detects the signal and reports attendance. Signal strength (RSSI) can be used to estimate distance.
Assessment: BLE answers the presence question well and the identity question not at all. Its most common weakness is not even an exotic attack: leaving the phone on the desk and walking out. That is why BLE must never count as evidence on its own, only as a layer that binds some other identity check to a place.
The student is required to be connected to the campus WiFi, to be attached to a particular access point (AP), or to arrive from a particular IP range.
Assessment: Moodle's Attendance plugin uses this elegantly: IP restriction plus a “block shared IPs” option. That is, more than one student cannot check in from the same IP address — a simple, effective measure against bulk attendance entered from a single device. It is notable that a free open-source plugin offers a better anti-fraud idea than many paid products.
The student's GPS position is checked against the campus or building polygon.
Assessment: Location data is one of the signals with the lowest security value and the highest privacy cost — a bad trade. Security should not be built on it; at most it should be used as a weak inconsistency indicator.
The student presses a finger onto a reader; the template is matched against a central or on-device database.
A camera (fixed, or the teacher's phone) scans the room. Every face in the frame is detected, its template is extracted and compared against the template database of the whole class (sometimes the whole school). This is called 1:N recognition.
The quality of face recognition depends not on distance but on how many pixels the face occupies. On a typical phone camera (26 mm equivalent, horizontal field of view ~69°) the arithmetic works out like this:
| Distance | Face width at 720p | 1080p | 4K |
|---|---|---|---|
| 2 m | 85 px | 127 px | 254 px |
| 4 m | 42 px | 63 px | 127 px |
| 6 m | 28 px | 42 px | 85 px |
| 8 m | 21 px | 32 px | 63 px |
The inter-ocular distance (IOD) the standards require, meanwhile: the lowest mobile profile of ANSI/NIST-ITL asks for 90 pixels, the Frontex border-crossing guidelines for 90 pixels, and the ICAO passport standard for 120 pixels. The inter-ocular distance of a student six metres away at 1080p: 17 pixels.
The consequences of this have been measured cleanly in independent benchmarks:
And as the gallery grows, the probability of a false match rises linearly (FPIR ≈ N × FMR). NIST's demographic evaluation (NISTIR 8429) measured false-match rate differences between groups of up to 714-fold at the same threshold; NISTIR 8280 additionally found that false positives rise among children and the elderly.
Documented case: In December 2025 in Wenzhou, China, employees at a neighbourhood committee office printed their colleagues' faces, turned them into masks, fooled a face-recognition time-and-attendance system and clocked in without going to work. The press called it the “$2 hack”. Face recognition without liveness detection is broken by a piece of paper.
Legal position: In 2019 in Sweden, the municipality of Skellefteå ran a three-week face-recognition attendance pilot with 22 students at a secondary school. The outcome was Sweden's first GDPR fine. The school argued it had obtained consent; the data protection authority rejected this — there is a “clear imbalance of power” between a school and a student. The key sentence in the decision: “there is a legal basis for monitoring students' attendance, but there is no legal basis for doing so by processing special category data.” Furthermore, the European Data Protection Board's Guidelines 3/2019 treat the creation of momentary templates for people who have not consented but who enter the frame as unlawful — even if the template is deleted immediately.
The student takes a selfie on their own phone. The app compares the face against a single template belonging to that student alone. This is called 1:1 verification and it is a categorically different problem from 1:N recognition. The template never leaves the device's secure hardware (Secure Enclave / StrongBox).
Assessment: 1:1 verification solves the identity question but does not solve the presence question. Used on its own it turns into a system for checking in from home. Its value only appears when it is combined with a proximity layer. And because the security model rests entirely on device integrity, hardware-backed key attestation and app integrity verification are not optional but mandatory components.
| Method | Identity | Presence | Continuity | Privacy | Friction | Cost |
|---|---|---|---|---|---|---|
| Paper / signature | ✗ | ~ | ✗ | ✓✓ | High | Zero |
| ID card (RFID/NFC) | ✗ | ✓ | ✗ | ✓ | Medium | Hardware |
| Static QR | ✗ | ✗ | ✗ | ✓ | Low | Zero |
| Dynamic / animated QR | ✗ | ~ | ✗ | ✓ | Low | Zero |
| Code / PIN | ✗ | ✗ | ✗ | ✓ | Low | Zero |
| Clicker | ✗ | ✓ | ~ | ✓ | Medium | High |
| BLE proximity | ✗ | ✓ | ✓ | ✓ | Very low | Low |
| WiFi / IP | ✗ | ~ | ✓ | ~ | Very low | Zero |
| GPS / geofence | ✗ | ✗ | ~ | Weak | Very low | Zero |
| Fingerprint | ✓✓ | ✓ | ✗ | ⛔ Legal risk | High | Hardware |
| Face recognition (1:N room scan) | ~ Depends on distance | ✓ | ~ | ⛔ Legal risk | Low | Medium |
| Face verification (1:1 on device) | ✓✓ | ✗ | ✗ | ✓ | Low | Low |
What the table says is plain: no single method solves identity, presence and continuity at once. The methods that solve identity best (the biometric ones) do not solve presence, or run into a legal wall; the methods that solve presence best (BLE, cards) do not solve identity at all.
Indiana University's official academic integrity guidance reaches the same conclusion and states it openly: no single technological solution is flawless; multiple verification methods should be combined.
The conclusion we drew from the table above shaped yoklama.ai's architecture directly. We combine three layers — each one covering another's blind spot:
The student verifies on their own phone. The template stays in the Secure Enclave / StrongBox, never leaves the device, and no biometric field is ever sent to the server. Handing the phone to a friend does not work. Printed photographs and screen replay attacks are eliminated by a liveness model that works at full capacity in close-up.
Why not 1:N: The pixel arithmetic shows that room scanning does not work at a distance. 1:1 verification works at full capacity in close-up, and a template is created only for the student who chooses that route.
The teacher's device acts as an “anchor”: over BLE and as an animated QR code on screen, it broadcasts an unpredictable ticket renewed every 1.5 seconds. The student's verification proof is bound to that ticket by being signed with a hardware-bound, non-exportable key. A phone that has not seen the ticket cannot produce a valid proof; that is, checking in from home on your own is not possible. A friend inside the room forwarding the ticket outside (a relay attack), however, is not closed by this layer — only the teacher's spot check deters that, and only probabilistically.
Attendance is not tied to a single moment. Two or three rounds are run during class at times nobody knows in advance. The student cannot know when verification will be requested; the “check in and leave” trick does not work. A participation ratio is computed and the decision is made against the institution's own threshold.
In every round the system picks a few students at random and asks the teacher: “Is Ayşe Y. in the room?” This catches — probabilistically — the one attack cryptography cannot solve: the relay attack. A friend in the room can forward the ticket outside; but if the student outside comes up in a spot check, they are caught.
yoklama.ai is not unbreakable; no attendance system is. The relay attack has not been fully solved — the time window has only been narrowed and compensated for with human checking. Identical twins are a limit no face system can overcome, and it is not being hidden. On a fully compromised device, client-side verification is not absolute; in that case the system does not say “present”, it closes the untrusted channel and routes to the teacher-approved path.
The claim we make is this: to raise the cost of cheating from signing your name to a coordinated technical attack, and to leave a probability of being caught on every attempt.
The teacher chooses the method. For each class, the teacher who opens the session decides which routes count as valid; the institution administrator can narrow this menu and may never enable face verification at all. When proximity + device binding are enabled, a student who does not use face verification takes part at the same speed with no extra step. Only in a class where face verification alone is enabled is the alternative the teacher marking by hand. In addition, the student sees their own attendance record immediately and can appeal it — of the 18 competing products we examined, we found a structured appeal workflow in none.
To be honest: domestic student information systems such as Proliz and UBYS offer attendance inside their own products — a separate purchasing decision for us, a module for them. Moodle's Attendance plugin is free and offers rotating QR + IP restriction + shared-IP blocking. Jibble is free with unlimited users. SEAtS positions itself strategically above us by connecting attendance to a student risk engine with more than 50 signals. These are real disadvantages and there are answers to them on our roadmap.