Fingerprinting, PUFs and Quantum PUFs: How a Device Proves What It Is

In short

No two chips, cameras or radios come out of a factory exactly alike. Those tiny differences can identify a device, the way a fingerprint identifies a person.
A fingerprint is observed from the outside and is not secret. A PUF is built into the chip and answers questions only that chip can answer, so it can produce keys without storing them anywhere.
Classical PUFs have been broken by machine learning, by physical cloning and by reading their side effects. Quantum PUFs use the rule that quantum states cannot be copied, so an attacker cannot simply record and replay them.
Quantum PUFs are still mostly in the lab. They need cheap hardware, a way to cope with noise, and verifiers that do not need large quantum memories.

The problem: where does a device keep its secret?

Most device security rests on a secret key. A payment card, a car key, a smart meter and a server's root of trust all prove who they are by doing something with a key that no one else has. The weak point is where that key lives. It usually sits in non-volatile memory, and anything stored can in principle be read out. Attackers decap chips, probe buses, image fuses with electron microscopes, and glitch the power supply until the chip leaks. Protecting stored keys against that kind of attack takes shields, sensors and tamper meshes, and it costs money on every unit.

There is another option. Manufacturing is never perfect. Two transistors drawn identically on a mask end up with slightly different threshold voltages. Two radios from the same production line have slightly different oscillators. Two sheets of paper have different fibre patterns. These variations are random, cheap to get, and very hard to reproduce on purpose. The idea behind both fingerprinting and PUFs is to use them as identity: let the device be its own key.

Fingerprinting: identity from what a device already does

Device fingerprinting measures characteristics a device shows in normal operation and compares them with a stored profile. Nothing is added to the device. The verifier simply observes it closely enough to tell it apart from others.

Radio fingerprints. Every transmitter has small imperfections in its oscillator, amplifier and modulator. Brik and colleagues showed in 2008 that these imperfections identify individual Wi-Fi cards with high accuracy. Similar methods are used to spot cloned or spoofed transmitters in cellular, satellite and IoT networks.
Clock skew. Kohno, Broido and claffy showed in 2005 that a computer's clock drifts at its own rate, and that the drift can be measured remotely from TCP timestamps, even through a network.
Camera sensors. Each image sensor has a fixed pattern of pixel sensitivity, called PRNU. Lukáš, Fridrich and Goljan used it in 2006 to match photos to the camera that took them, and forensic labs use it today.
Browsers. Eckersley's 2010 study found that the combination of fonts, plugins, screen size and settings made most browsers in his sample unique. Advertising and fraud-detection systems rely on the same effect.

Fingerprinting is attractive because it needs no cooperation from the device and works on hardware already in the field. Its weakness follows from the same fact. A fingerprint is a property anyone can observe, so it is not a secret. If an attacker can measure it, they can record it and, with enough effort, play it back. A software-defined radio can be tuned to imitate another transmitter's quirks. A browser's reported settings can be faked. Fingerprints are good evidence of identity, and useful for forensics and for flagging anomalies, but on their own they are weak authentication against a prepared attacker. They also raise a privacy problem: something that identifies a device without its owner's consent can be used to track that owner.

PUFs: a challenge only one chip can answer

A physical unclonable function, or PUF, takes the same manufacturing randomness and turns it into a function. You send the device a challenge, and the device's physical structure determines the response. The mapping from challenges to responses is fixed for a given device, different for every device, and not something the manufacturer can choose or copy. Pappu, Recht, Taylor and Gershenfeld introduced the idea in Science in 2002 with an optical version: a laser shone through a block of epoxy filled with tiny glass spheres produces a speckle pattern that changes completely if the beam moves slightly. The same year Gassend, Clarke, van Dijk and Devadas built the first silicon PUFs from circuit delays.

Several designs are now common:

SRAM PUFs. When static memory powers up, each cell settles to 0 or 1 depending on which of its two transistors is slightly stronger. The power-up pattern is random across chips and mostly stable on one chip. Guajardo, Kumar, Schrijen and Tuyls showed in 2007 that it works on ordinary FPGAs, and it is the most widely deployed PUF today because almost every chip already has SRAM.
Ring-oscillator PUFs. Many identical oscillator loops run at slightly different frequencies. Comparing pairs gives bits. Suh and Devadas described the design in 2007.
Arbiter PUFs. A signal races down two paths whose route is set by the challenge bits, and a latch at the end records which path won. Every challenge configures a different race, so the number of challenges grows exponentially with the number of stages.
Optical and paper PUFs. Scattering materials, fibre patterns and printed random features, often used for anti-counterfeiting labels rather than chips.

PUFs split into two kinds. A weak PUF, such as SRAM, has only a few challenges and is used inside the chip to produce one key. A strong PUF, such as an arbiter PUF, has so many challenges that a verifier can use a fresh one each time and never repeat it. The names describe the number of challenges, not the level of security.

How fingerprints and PUFs differ

Both use the same raw material, uncontrolled physical variation, but they are built for different jobs. The table puts them side by side with the quantum PUFs discussed further down.

	Fingerprint	Classical PUF	Quantum PUF
Where it comes from	Side effects of normal operation	A circuit or material designed to be random	Quantum states sent to, or produced by, a random physical system
How it is read	Passively, by observing the device	Actively, by sending a challenge	Actively, with challenges or responses that are quantum states
Is it secret?	No. Anyone close enough can measure it	The responses are, until someone models or reads them	The challenge cannot be fully learned, even by someone who intercepts it
Can it make a key?	Not reliably	Yes, with error correction	Yes in principle; noise handling is still being worked out
Main attacks	Recording and replay, imitation	Machine-learning models, side channels, physical cloning	Mostly open; the theory shows limits as well as guarantees
Maturity	Widely used in forensics and fraud detection	In commercial FPGAs and secure chips	Lab demonstrations and security proofs

Fingerprints identify. PUFs authenticate and generate keys. Quantum PUFs aim to do the same without trusting that no one is listening.

The short version: a fingerprint is something a device has, and a PUF is something a device does. A fingerprint answers the same way to everyone. A PUF answers a question you chose, and its answer is useless to an eavesdropper if you never ask that question again.

How PUFs provide security

PUFs are used in two main ways.

Authentication

EnrolVerifier sends many challenges and stores the response pairs
ChallengeIn the field, verifier picks a pair it has never used
RespondDevice answers from its physical structure
CheckClose enough to the stored response? Accept, and discard that pair

Key generation

EnrolChip reads its PUF once and writes public helper data
Power upChip reads the PUF again; a few bits have flipped
CorrectHelper data fixes the flipped bits
DeriveThe same key every time, held only while the chip is on

In both uses the secret is the physical structure itself. Nothing secret is written to memory.

Authentication uses a strong PUF. During enrolment, in a trusted setting, the verifier collects a large set of challenge-response pairs. Later, to check a device, it sends a challenge from that list and compares the response. Each pair is used once, so an eavesdropper who records the exchange learns nothing about the next one.

Key generation uses a weak PUF. The difficulty is that PUF readings are noisy: a few percent of SRAM bits come up differently on each power-up, and more at temperature extremes or as the chip ages. A key that changes is useless. The fix is a fuzzy extractor, formalised by Dodis, Reyzin and Smith in 2004. At enrolment it stores helper data that lets the chip correct its own read-out errors later, using an error-correcting code. The helper data can be public. The chip reconstructs the key at power-up, uses it, and forgets it at power-down.

This gives three security properties that stored keys lack. There is no key at rest to extract with the power off. Invasive attacks tend to change the physical structure and destroy the very thing they try to read. And the manufacturer never knows the key, so a compromised supply chain cannot leak a database of them.

PUFs also fit well with the move to post-quantum cryptography. ML-KEM and ML-DSA, the NIST standards from 2024, derive their key pairs from short random seeds of 32 bytes. A weak PUF can supply that seed, so a chip can hold a post-quantum identity key that is rebuilt at each boot and never stored. The PUF protects the key at rest. The post-quantum algorithm protects it in use against a future quantum computer.

How PUFs get broken

The word “unclonable” is a design goal, not a guarantee. Twenty years of attacks have shown where real PUFs fall short.

Machine-learning models. Rührmair and colleagues showed in 2010 that an arbiter PUF behaves almost like a linear function of its challenge bits. After observing a few thousand pairs, logistic regression predicts its responses with better than 99% accuracy. XOR-ing several arbiter chains raised the cost but did not remove the risk.
Reliability attacks. Becker showed in 2015 that the PUF's noise leaks information: challenges whose responses sometimes flip sit close to a decision boundary. Measuring which ones are unstable let him model XOR arbiter PUFs, including ones in commercial RFID tags, that resisted the earlier attacks.
Physical cloning. Helfmeier, Boit, Nedospasov and Seifert read the power-up state of an SRAM PUF through the back of the chip in 2013, then used a focused ion beam to edit a second chip so it produced the same pattern. It took expensive equipment, but it ended the idea that PUFs cannot be copied at all.
Side channels. Power draw, timing and electromagnetic emissions during the read-out and the error correction can leak response bits, just as they leak keys from conventional cryptography.
Helper-data manipulation. If an attacker can rewrite the public helper data and watch whether the chip still works, each attempt can reveal a bit of the key. Helper data needs integrity protection even though it is not secret.
Emulation of the reader. Many PUF schemes assume the verifier talks to the real physical object. An attacker who places a small computer between the reader and the device can answer with a model, and a classical reader cannot tell the difference.

That last point matters most for what comes next. A classical PUF's challenge and response are ordinary bits. Anyone on the channel can copy them, store them, and feed them to a model. Security then depends on the attacker not having enough pairs, enough compute, or close enough physical access. Rührmair and van Dijk argued in 2013 that many published PUF protocols fail under realistic attack models for exactly this reason.

Quantum PUFs

A quantum PUF brings quantum physics into the challenge, the response, or the device itself. The key tool is the no-cloning theorem: an unknown quantum state cannot be copied. If the challenge is a quantum state, an attacker who intercepts it cannot keep a copy, and measuring it to learn what it was disturbs it. The term covers four rather different approaches.

1. Quantum readout of a classical PUF

The device is still a classical random object, such as a scattering material, but the verifier probes it with quantum light. Škorić proposed this in 2012. Goorden, Horstmann, Mosk, Škorić and Pinkse demonstrated it in Optica in 2014: they shaped the wavefront of a light pulse containing fewer photons than the key has spatial modes, and checked that the light was focused correctly after scattering. With so few photons, an attacker cannot work out which wavefront was sent, so they cannot compute the right answer, even if the full structure of the key is public. This is the property classical PUFs lack: the reader no longer needs to trust that it is talking to the real object. Nikolopoulos and Diamanti extended the scheme to continuous-variable light in 2017, which suits standard telecom components.

2. Quantum effects with a classical readout

Here the randomness comes from quantum behaviour inside the device, but the challenge and response are classical. Roberts and colleagues reported in 2015 that resonant tunnelling diodes give a unique current-voltage signature set by atomic-scale variations, readable at room temperature. These designs are hard to clone physically, but their outputs are still ordinary data once read, so the modelling and replay concerns from the classical side remain.

3. Fully quantum PUFs

In the most ambitious version the device applies an unknown unitary transformation to whatever quantum state it receives. The challenge is a quantum state and so is the response. Arapinis, Delavar, Doosti and Kashefi gave the first formal security analysis in Quantum in 2021. They proved limits as well as guarantees. No quantum PUF can stop an attacker from producing some valid challenge-response pair after enough access, so the strongest notions of unforgeability are impossible. A unitary quantum PUF can stop an attacker from answering a challenge the verifier picks at random, which is the property authentication protocols need. Doosti, Kumar, Delavar and Kashefi then built identification protocols on that guarantee, with security proofs that hold against attackers who have quantum computers.

4. Fingerprinting quantum computers

The last approach turns the idea around. Every quantum processor has its own pattern of qubit frequencies, error rates and crosstalk. Phalak and colleagues used this in 2021 to build quantum PUFs on cloud quantum computers. On IBM hardware, responses from different machines differed in about 55% of bits, while repeated runs on the same machine differed in about 4%. That matters for anyone renting quantum time: a user may want proof that a job ran on the machine they paid for and not on a cheaper or simulated one.

What quantum PUFs improve

Interception stops being useful. With classical PUFs, recording the traffic gives the attacker training data. A quantum challenge cannot be copied, and measuring it disturbs it, so the record is incomplete and the disturbance can be detected.
Emulation becomes much harder. To fake a response, an attacker has to know the challenge. If the challenge is hidden by quantum mechanics, a model of the device is not enough, even a perfect one. This removes the trusted-reader assumption that many classical schemes quietly depend on.
Security rests on physics, not on the attacker's budget. Machine learning attacks on classical PUFs get better with more data and compute. A no-cloning argument does not.
A path to formal proofs. The quantum PUF literature has defined precisely what unforgeability means and which versions can be reached. Classical PUFs were deployed long before such definitions existed.
Identity for quantum hardware and networks. Nodes in a quantum network and machines in a quantum cloud need to be authenticated too. A quantum PUF can do that using the same channels and hardware the network already uses.

What remains open

Quantum PUFs are promising on paper and work in careful lab setups. Much has to change before they replace the SRAM PUF in a microcontroller.

Building a good unitary cheaply

The security proofs assume the device behaves like a Haar-random unitary, which would need an exponentially large circuit. The practical substitute is a pseudorandom unitary. Ma and Huang showed in 2024 how to build these from any quantum-secure one-way function, which settles the question in theory. What is missing is a physical device, shaped by manufacturing randomness rather than a programmed key, that provably behaves like one.

The verifier's quantum memory

If the responses are quantum states, the verifier has to store them or regenerate them to compare. Storing many quantum states for months is far beyond current quantum memories. Some protocols reduce the burden by letting the verifier use classical checks or shifting the quantum work to one side, but each trade-off weakens or complicates the security argument.

Noise, loss and error correction

Classical PUFs solved noise with fuzzy extractors. The quantum equivalent is not settled. Photons get lost, detectors miss counts, and the device itself drifts. Comparing quantum states uses tests such as the swap test, which only gives a statistical answer, so the verifier needs many rounds and a threshold that tolerates honest noise without letting forgeries through.

Hardware cost and size

The appeal of an SRAM PUF is that it costs almost nothing. Optical quantum readout needs a controlled light source, wavefront shaping and sensitive detectors. That is acceptable for a high-value credential, a vault or a network node, and far too much for a sensor tag. Integrated photonics may close part of the gap.

Side channels in quantum hardware

Quantum key distribution taught the field that real detectors and sources leak information that ideal models leave out. Quantum PUF readers will have the same problem. Very little work so far has attacked an actual quantum PUF setup the way Becker and Helfmeier attacked classical ones.

Better models for classical PUFs

Classical PUFs will remain the workhorse for years, and they still lack agreed security definitions. It is hard to say how many pairs an attacker may collect, or how much entropy an SRAM pattern really has after accounting for spatial correlation and ageing. The ISO/IEC 20897 series, with a test and evaluation part published in 2022, is a start, but certification practice is still young.

Privacy

Anything that identifies a device can track its owner. Fingerprinting already does this on the web and over radio. PUF-based protocols need ways to prove a device is genuine without revealing which device it is, for example through group signatures or anonymous credentials, and those add their own cost.

Fitting into real systems

A PUF is one component. It has to feed a key into secure boot, attestation and post-quantum algorithms, survive firmware updates, and handle the day a device is repaired, recycled or reported stolen. These engineering questions decide whether a design is used, and they get far less research attention than the physics.

What to do now

For most organisations, quantum PUFs are something to watch, not to buy. The useful steps today are classical. Find out where device keys are stored, and whether the hardware you already have includes a PUF-based key store. Check that the keys it protects are moving to ML-KEM and ML-DSA. Treat fingerprints as one signal among several, not as authentication. If you run quantum hardware or a quantum network testbed, start thinking about how its nodes will prove who they are, because that question will arrive with the first shared deployments.

Fingerprints, PUFs and quantum PUFs: how a device proves what it is.