The FFIEC's authentication guidance of 11 August 2021 lists device identification by a "complex digital 'fingerprint'" among its example controls. Device cookies, geolocation and IP address matching fare worse. The guidance calls them "insecure and ineffective if used alone", though they can be combined with other controls. An examiner reviewing your fraud controls will ask what the device check adds, and what happens when it fails.
What is device fingerprinting?
Device fingerprinting recognizes a device by what it reveals about itself, not by a tag placed on it. The W3C's fingerprinting guidance (Group Note, 25 September 2025) defines the web form, browser fingerprinting, as the capability of a site to identify a visiting user, user agent or device "via configuration settings or other observable characteristics".
Fraud teams want it because nothing has to be stored on the device: a fraudster who clears cookies keeps the same screen, fonts and graphics card. The guidance sees both sides: fingerprinting can serve as "a security measure" and is also "a threat to users' privacy".
Device fingerprinting, browser fingerprinting and cookies
Browser fingerprinting is the part a web page can read. Passive fingerprinting uses what arrives in web requests, "without the use of any code executed on the client"; active fingerprinting runs a script to read more.
Apps see more. Their platform identifiers, though, are scoped and resettable. On Android 8.0 and later, ANDROID_ID is unique per signing key, user and device, and may change after a factory reset. Apple's identifierForVendor changes once all of one vendor's apps are deleted and one is reinstalled. A cookie is stored on the device, and the user can delete it. A fingerprint is derived again on every visit.
How does device fingerprinting work?
A script or app SDK gathers attributes, condenses them and compares the result with devices seen before. The output is a likelihood.
- Collect. Passive attributes come with every request: headers, IP address, the TLS handshake. Active ones need code on the device: canvas and audio rendering, installed fonts, screen size, the number of processor threads.
- Hash. The values become one short identifier.
- Match. An identical hash means the same configuration. A similarity score across attributes means probably the same device.
- Allow for change. In the EFF's 2010 Panopticlick study, 37.4 percent of 8,833 users who accepted cookies and came back over more than 24 hours changed fingerprint at least once. A simple heuristic still guessed which earlier fingerprint a changed one had grown from, with 99.1 percent of guesses right and 0.86 percent false positives. Uniqueness depends on the crowd: 83.6 percent of 470,161 browsers in that privacy conscious sample were unique, against 33.6 percent of 2,067,942 fingerprints from one of the top 15 French websites in the 2018 study Hiding in the Crowd.
An illustrative case: a device shows 12 attributes and a browser update changes 2. An exact hash calls it a new device; a 10 of 12 similarity calls it the same one. California's privacy statute calls identifying a consumer or their device from personal information "to a degree of certainty of more probable than not" a "probabilistic identifier".
Where does device fingerprinting fail?
| Cause | What changes | What a fraud team sees |
|---|---|---|
| Tor Browser | Users made alike, in few "buckets" per attribute | Many unrelated users on one fingerprint |
| Brave (design of 2020) | Some values randomized per site and per session | One device, a new fingerprint every session |
| Safari 26.0 (September 2025) | Known fingerprinting scripts lose screen, canvas, audio and hardware readings | Fewer attributes from scripts Safari classes as known |
| Firefox 145 (November 2025) | Private Browsing and strict mode protections; unique users cut by almost half, according to Mozilla | Fewer unique fingerprints from those users |
| Emulators and automation tools | Software answers the script instead of real hardware | Attribute sets no real device shows |
| Factory reset or app reinstall | ANDROID_ID may change on reset; identifierForVendor on reinstall | A known device returns as new |
| Identical phones | Same model, same settings, same values | Many owners on one fingerprint |
| iOS apps | Apple forbids fingerprinting, tracking permission or not | Only the signals Apple allows |
Where device fingerprinting fails (as of November 2026).
Browsers reduce fingerprinting without removing it: W3C guidance calls complete elimination by a determined adversary, through widely deployed technical means alone, "implausible". A faked attribute is itself a signal, the core of spoofing detection: a desktop font list on a phone, or a time zone at odds with the IP address. Identical phones are harder still: in the 2018 study only 18.5 percent of 251,166 mobile fingerprints were unique.
Where does device fingerprinting sit in fraud and AML work?
Fraud teams run it at onboarding, login and payment. Payments law already names the device. In the EU, the transaction monitoring factors in Article 2 of the strong customer authentication standards name the access device only when the payment service provider supplied it.
The Payment Services Regulation text the Council adopted, not yet law, adds "device data, including identifiers of the device from which the transaction was initiated" to the payer's provider's fraud monitoring (Article 87(4)(e)). Within sharing arrangements, a provider with objectively justified reasons to suspect a user of fraud would share that data to the extent necessary. The payer's behavioral characteristics stay out of the exchange.
In the US, FinCEN's 314(b) fact sheet lists "device identification numbers" among the information institutions could share under the 314(b) safe harbor.
complead, formerly Sanction Scanner, counts device fingerprinting among the first party signals of its fraud prevention, beside network signals, behavioral biometrics, and emulator and tampering detection. The device and behavior analysis watches each session, and the device behind it, from the step before login up to the payment. Emulators, automation tools, spoofing tools and tampered runtimes each get a separate flag, one device behind several accounts shows up, and every score shows its drivers. It records no keystrokes and does not gather fingerprints in the biometric sense, face data or voice data.
What do fraud teams use device fingerprinting for?
Picture an illustrative sportsbook on a derby weekend. Forty new accounts claim the same welcome bonus, and 37 of them sign up from 3 devices. Every name differs and every document checks out, so only the devices link them.
Account takeover
Here the signal runs the other way: a known account appears on a device it has never used, for example just before a payee change, one route in account takeover fraud. NIST's digital identity guidelines let fraud indicators lower the risk of misauthentication but say they do not "substitute for an authentication factor". So a new device should trigger a step up, such as a fresh one time code. It should never log anyone in.
Multi accounting, bonus abuse and device farms
Back at the sportsbook, three devices behind 37 accounts need an analyst's review. The same NIST guidelines, written for devices registered as authenticators, say that, in applications open to the public, providers "SHOULD NOT prevent a device from being registered as an authenticator by multiple subscribers". They may still set limits against fraud at scale, such as on "the total number of subscriber accounts a single device can be registered with". A household shares a tablet. One bettor running 12 accounts is the case a device limit is for. Device farms, racks of real or emulated phones, rotate attributes so each account looks new, which is why application fraud checks weigh the device beside identity data.
Which privacy rules apply to device fingerprinting?
In the EU and the UK the rule is about access to the device, whether or not the data is personal. Article 5(3) of the ePrivacy Directive requires consent to store or read information on a user's device, and the EDPB's Guidelines 2/2023 (version 2.0, 7 October 2024) confirm that fingerprinting falls within it.
The UK's PECR regulation 6, as substituted on 5 February 2026, covers "collecting or monitoring information automatically emitted" by the device. The In the EU, the GDPR governs the result where an online identifier makes a person identifiable.
| Rule | Where | What it covers | Fraud exemption | Status |
|---|---|---|---|---|
| ePrivacy Directive Article 5(3), EDPB Guidelines 2/2023 | EU | Storing or reading information on a device, fingerprinting included | Only what is strictly necessary for a service the user explicitly requested | In force |
| GDPR Article 4(1), recitals 30 and 47 | EU | Online identifiers identifying a person | Legitimate interest test, for strictly necessary fraud prevention | In force |
| PECR regulation 6 and Schedule A1 | UK | Device access, including information the device emits | Strictly necessary to prevent or detect fraud in the service requested | In force since 5 February 2026 |
| CCPA section 1798.140 | California | Consumers' device identifiers, persistent or probabilistic | Security and integrity, a business purpose if reasonably necessary and proportionate | In force |
| Digital Omnibus, proposed GDPR Article 88a | EU | Would add a device access rule for personal data to the GDPR | Securing the controller's requested service, where necessary | Proposed, awaiting committee decision |
| Payment Services Regulation Article 87(4)(e) | EU | Device data in the payer's provider's fraud monitoring | Processing "in so far as that is necessary" | Awaiting Parliament's second reading |
Which rules apply to device fingerprinting (as of 15 November 2026).
Does a fraud fingerprint need consent?
In the EU, yes, unless reading the device is strictly necessary for a service the user explicitly requested. The Article 29 Working Party's Opinion 9/2014, guidance rather than law, extends the security exemption to device fingerprinting, but not to the security of websites or "third party services" the user did not explicitly request, and data collected under it "may not be used for any secondary purposes". The EDPB leaves the exemptions to case by case analysis under each member state's law and its authorities' guidance. In the UK, Schedule A1 paragraph 4(2)(c) exempts access strictly necessary "to prevent or detect fraud in connection with the provision of the service requested".
Where the script runs decides the answer. On login and payment pages, a fraud fingerprint protects the service the customer asked for. The same script on marketing pages, or one whose output also feeds ad attribution, falls back under consent. GDPR recital 47 guides the next step, processing the result.
Which device fingerprinting rules moved in 2026?
- 19 January. The Digital Omnibus went to Parliament's committees; its procedure file) still read "awaiting committee decision" on 15 November.
- 5 February. UK PECR regulation 6 and Schedule A1, with the fraud exemption, came into force.
- 12 June. FinCEN's new 314(b) fact sheet named device identification numbers.
- 29 October. Device data under Article 87(4)(e) went into the Payment Services Regulation text the Council adopted; the second reading) has an indicative plenary date of 14 December 2026.
One decision is left for your firm: which pages run the fingerprint script and for what purpose, agreed in writing by fraud and privacy before launch. In the EU that line decides whether consent is needed; in the UK, whether the fraud exemption applies.
Related terms
- Account takeover fraud. Someone takes over a real customer's account and acts as that customer.
- Spoofing. Disguising a message, call or website as coming from a trusted source.
- Identity fraud. Someone else's personal details used to obtain money, goods or credit.
- Synthetic identity fraud. An invented person built from real and made up details.
- Chargeback fraud. A cardholder disputes a purchase the cardholder really made.
- Fraud score. One number for an event's risk, built from signals such as device, identity and payment.