Firefox 143 cuts uniquely identifiable users from 75% to 46%, Mozilla study says

Mozilla Corporation, the University of Toronto and UC San Diego have published a large study of browser fingerprinting, with a companion site at fingerprinting.study. The dataset is telemetry from almost 25 million Firefox release users. The team used it to measure how identifiable browsers really are, to find which characteristics make them unique, and to design the fingerprinting protections that shipped in Firefox 143. The authors say those protections reduce the fraction of uniquely identifiable users from 75% to 46% while maintaining web compatibility.

The central finding is that uniqueness comes from a small set of high-impact characteristics and their combinations, not from piling up many independent measurements. The authors compare it to the 80/20 rule. For Fingerprint Pro, which they call the most discerning fingerprinter they model, 80% of the users it sees as unique are already unique after only 8 metrics, though it collects over 50 in total. Signifyd collects over 80 metrics but identifies fewer users as unique, and 80% of the users it sees as unique are identified after 9 metrics.

The authors argue that a single headline percentage of unique users is misleading, because it mixes populations that behave differently. The operating system is always exposed to a fingerprinter (lying about it breaks shortcut keys and gives wrong downloads), so the study segments users by OS. Each OS has different important metrics. Clamping the available screen resolution, for example, takes 25.3 percentage points off cumulative macOS uniqueness but between 0 and 3 points off the other three platforms. Averaged across all users, that 25-point drop shrinks to 3.5.

Two more methodological points. First, high Shannon entropy of a metric does not reliably predict its fingerprinting impact. Second, earlier work counted only unique users, not the rest. The study uses the Simpson Index (also called the Herfindahl-Hirschman Index) to measure cohort structure: whether non-unique users sit in one huge group of 1 million or in 500,000 groups of two. This matters when ranking defenses. Two protections can each lower the share of unique users by about 2.1 percentage points, yet one improves the cohort score by 56% and the other by only 13%.

On defenses, the authors say standardization (making users report the same values) makes users look more alike but never makes everyone look the same, and the problem is worse on sites with small visitor populations. It should be paired with randomization, which is very strong when it works but can always be detected and bypassed. Neither can carry the privacy burden alone. Correlating fingerprint values with hardware and other traits let the team find why values differ between users; addressing those root causes has already produced more consistent outputs for image rendering, Web Audio and mathematical operations on Android, with more causes under investigation.

Compatibility shaped what shipped. The authors note that breakage from a strict fingerprinting mode arrives later and users do not connect it to the setting they enabled, so people post for help confused about why a site fails in Firefox. They treat compatibility as a precondition for a default, not a trade-off against privacy, and say the hard part is finding what breaks on a vast web.

The data came from the Firefox user-characteristics ping, a dedicated telemetry ping that must be explicitly enabled (which has happened twice, once in 2025 and once in 2026) and sends one report before going dormant. It carries no client ID, only a per-ping UUID used to de-duplicate users, and is sent over Oblivious HTTP through a relay operated independently of Mozilla, so the relay sees the IP address but not the contents and Mozilla sees the contents but not the IP address. Access was restricted to four authorized Mozilla employees, and the data was retained for 90 days and has since been deleted.

The site offers the paper (an extended PDF), full appendix data with every rank for each operating system and all three modeled services, aggregate data and analysis code, a discussion board for follow-up research, and an interactive test that runs entirely in the visitor's browser and compares its measurements with the published population data; no data is stored or submitted to the authors. The authors say they deliberately did not publish everything the dataset supports, because commercial fingerprinters borrow from the research community and they would rather leave a question open than make it easier to build a better fingerprinter.

Key facts

• The study uses telemetry from almost 25 million Firefox release users; protections designed from it shipped in Firefox 143 and, per the authors, cut the share of uniquely identifiable users from 75% to 46% while maintaining web compatibility.
• A small set of metrics drives uniqueness: for Fingerprint Pro, 80% of the users it sees as unique are unique after 8 of its 50-plus metrics; for Signifyd, 80% are identified after 9 of its 80-plus.
• Numbers differ sharply by operating system: clamping screen resolution removes 25.3 percentage points of macOS uniqueness but 0 to 3 points on the other three platforms, and 3.5 points averaged across all users.
• The authors score cohort structure with the Simpson Index: two protections that each cut unique users by about 2.1 percentage points improved the cohort score by 56% and 13% respectively.
• Telemetry was sent over OHTTP with no client ID, seen by four authorized Mozilla employees, kept for 90 days and since deleted.

Why it matters

Browser fingerprinting is a known tracking mechanism, but the authors say its real-world effectiveness has been hard to measure. This study puts numbers on it from a population of almost 25 million Firefox release users and ties them to protections that actually shipped in a mainstream browser, Firefox 143. The reported drop from 75% to 46% uniquely identifiable users is a share of users, not a relative change. The findings also challenge common habits: single headline percentages, Shannon entropy as the yardstick for a metric, and judging a defense by uniqueness alone.

Who it affects

Firefox users are the direct subjects and beneficiaries, since the protections shipped in Firefox 143. Browser vendors and privacy researchers get a method for ranking defenses by measured impact and cohort structure. Commercial fingerprinters such as Fingerprint Pro and Signifyd were modeled in the study. Users of different operating systems are affected differently, and the authors stress that per-OS breakdowns matter more than any single figure.

How to use it

Visitors can run the interactive test on the companion site, which shows their own value for each characteristic and how many of the roughly 25 million measured browsers share it, overall and within their operating system. It runs entirely in the browser against static reference files. The site also offers the extended paper PDF, full appendix data, aggregate data and analysis code for download, and a discussion board for proposing follow-up research. Problems with the test can be reported as a GitHub issue, with the diagnostics from its Copy diagnostics button.

How solid is it

The study comes from Mozilla, the University of Toronto and UC San Diego, and the dataset is large. The 75% to 46% figure is the authors' own claim, taken from the paper abstract and site intro. It is a single aggregate number; the per-OS breakdown is in the paper and appendix data. The companion site is the source here, and its text names no individual authors. The authors also say they held back some analyses and data on purpose.

Risks and caveats

The 75% and 46% figures are aggregates, and the authors themselves warn that single percentages hide large differences between operating systems. The screen-resolution example is given to show per-platform variation, and the page does not say whether it is part of the Firefox 143 set. The page does not list which protections make up the Firefox 143 set or whether they are on by default for all users. The authors note that standardization never makes everyone look the same, that randomization can always be detected and bypassed, and that strict modes cause delayed breakage users do not trace back to the setting. Fingerprinters also evolve and borrow from public research.

“we would rather leave a question open than make it easier to build a better fingerprinter”

— The study's authors, fingerprinting.study