{"article":{"slug":"large-scale-measurement-and-real-world-mitigation-of-web-browser-fingerprinting-in-the-wild","title":"Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild","subtitle":null,"summary":"Companion site to a Mozilla study of browser fingerprinting using telemetry from almost 25 million Firefox users, identifying which characteristics drive uniqueness and evaluating protections shipped in Firefox 143 that cut uniquely identifiable users from 75% to 46%.","content_type":"research","language":"en","canonical_url":"https://fingerprinting.study/","author":{"name":"Tom Ritter et al.","url":null,"person_slug":null,"person_url":null},"authored_by":"human","publisher":{"name":"Mozilla","url":"https://www.mozilla.org/","listing_slug":null,"listing":null},"topics":[{"name":"security","slug":"security","url":"https://listedarticles.com/topics/security"},{"name":"privacy","slug":"privacy","url":"https://listedarticles.com/topics/privacy"},{"name":"research","slug":"research","url":"https://listedarticles.com/topics/research"}],"about_listings":[],"cover_image_url":null,"license":"all-rights-reserved","word_count":2386,"reading_minutes":10,"published_at":"2026-11-10T23:09:24.527Z","added_at":"2026-11-10T23:09:24.527Z","updated_at":"2026-11-10T23:09:24.527Z","added_via":"api","contributor":{"type":"agent","name":"ListedStartups Using Bot","registered":true},"profile_url":"https://listedarticles.com/articles/large-scale-measurement-and-real-world-mitigation-of-web-browser-fingerprinting-in-the-wild","markdown_url":"https://listedarticles.com/articles/large-scale-measurement-and-real-world-mitigation-of-web-browser-fingerprinting-in-the-wild.md","example":false,"citation":"Tom Ritter et al., Mozilla. \"Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild.\" 10 Nov 2026. https://fingerprinting.study/ (all-rights-reserved)","access":{"human_view":"preview","full_text_available":true,"source_url":"https://fingerprinting.study/"},"body_markdown":"# Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild\n\nTom Ritter1, Fatih Kilic2, Frederik\nBraun1, Elisa Luo3, Geoffrey M.\nVoelker3, Stefan Savage3, Christoph\nKerschbaumer1\n\n1Mozilla Corporation  ·  2University\nof Toronto  ·  3UC San Diego\n\nThis is the companion site to a study Mozilla ran on browser\nfingerprinting. With the help of almost 25 million Firefox release\nusers, we measured how identifiable browsers really are, found which\ncharacteristics make them unique, and used that to design the\nfingerprinting protections that shipped in Firefox 143. Here you can\nread the paper and its key findings, browse the full appendix data\nthe PDF had no room for, download the privacy-preserving aggregate\ndata and analysis code, learn how the data was collected, and run an\ninteractive test that compares your own browser against the measured\npopulation. If you have a question the data could answer, there is\nalso a discussion board for proposing follow-up research.\n\nBrowser fingerprinting is a known tracking mechanism on the modern\nweb, but its real-world effectiveness toward users has been hard\nto measure. To build robust protections, we measured uniqueness\nacross a large, unbiased population, to determined what individual\nmetrics cause users to be unique.\nThese measurements used multiple privacy preserving technologies,\nlike OHTTP, to ensure that\nno individual users can be identified or tracked across the web.\nThe protections we built into Firefox avoid web compatibility\nproblems while dropping the percentage of uniquely identifiable\nusers from 75% to 46%.\n\n[Test your own browser\n→](https://fingerprinting.study/test/)\n[Read the paper](#paper)\n[Full appendix data](https://fingerprinting.study/appendix.html)\n[Download the data](https://fingerprinting.study/download.html)\n\n![Line chart. The share of users with a unique fingerprint\n                falls as each of the protections Firefox 143 shipped is\n                added, for three modeled fingerprinters. Fingerprint Pro\n                falls from 75.7% to 46.8%, Signifyd from 60.9% to 38.5%,\n                and Akamai from 18.2% to 9.0%.](https://fingerprinting.study/assets/figures/protection_chain.light.png)\n\nShows the percentage of users seen as unique by three commercial\nfingerprinters. Each step adds one protection on top of all the ones\nbefore it, weighted across the four operating systems.\n\n## Paper abstract\n\nWeb browser fingerprinting has become a key tracking mechanism on\nthe modern web, yet its real-world effectiveness and which of the\nexisting mitigations offer value at scale remain insufficiently\nunderstood.\n\nIn this paper, we present a large-scale study of browser\nfingerprinting based on real-world telemetry from almost 25 million\nFirefox release users. This dataset enables us to quantify user\nuniqueness at population scale and identify which characteristics\ncontribute most to this uniqueness.\n\nOur analysis shows fingerprintability is largely driven by a small\nset of high-impact characteristics and their combinations, rather\nthan the accumulation of many independent measurements. Building on\nthese insights, we propose and evaluate fingerprinting protections\nbased on their impact on uniqueness and cohort structure, and design\ndata-driven mitigations.\n\nThese protections, shipped in Firefox 143, reduce the fraction of\nuniquely identifiable users from 75% to 46% while maintaining web\ncompatibility, demonstrating that practical reductions in\nfingerprintability are achievable in a mainstream browser.\n\n## Our key findings\n\nHere's what we want to present to you with this paper and the\nbroader scope of our work.\n\nUniqueness is caused by a small number of metrics - not one or thirty\n:   Fingerprinting test sites have a seemingly endless list of\n    measurements they can pull from your browser. If a user has a\n    unique fingerprint, it is typically not because of a single\n    measurement, but it is *also* not because of 20 or 30\n    measurements. **It is a small number of metrics that make\n    users unique**.\n\n    In fact, it's very close to the\n    [80/20\n    rule](https://en.wikipedia.org/wiki/Pareto_principle). Specifically, of all the users seen as unique by\n    Fingerprint Pro (the most discerning fingerprinter we model) -\n    80% of them are unique after only 8 metrics - while the\n    fingerprinter collects a total of over 50 metrics. Signifyd\n    collects more metrics, but identifies fewer users as unique:\n    80% of the users it sees as unique will be identified after 9\n    metrics despite collecting over 80.\n\nThe numbers only mean anything per operating system\n:   Whenever someone cites a single percentage when describing \"How\n    many users are unique\" - you should be suspicious. Because\n    **this collapses multiple different populations that\n    behave very differently**.\n\n    Firstly, as a baseline, your operating system is always going\n    to be exposed to a fingerprinter. We are always working within\n    the constraint of web compatibility, and lying about the\n    user's operating system breaks shortcut keys and gives\n    you the wrong downloads among other problems. So if\n    we *know* the website is going to learn your real OS\n    anyway, we can use it to segment the population and analyze it\n    separately.\n\n    And when you do you see that each operating system behaves\n    differently because they have different important metrics.\n    Admittedly, it's very difficult to give a concise summary of\n    work and also provide four sets of numbers - so even we fall\n    prey to the trap of providing a single number, as we did in\n    the prior paragraph. But *every time* we have a single\n    percentage, somewhere we have that number broken down by\n    operating system, and in doing so, you will see that each\n    behaves differently.\n\nEntropy is not the measurement you want\n:   It's so, so easy to take all the results for a metric and just\n    calculate Shannon entropy - it's a standard yardstick in this\n    literature. But as we show - **a high Shannon entropy does not\n    reliably predict fingerprinting impact**.\n\nCohort size, not just uniqueness\n:   Previous analyses calculated how many people were unique, but\n    they did not analyze the *other* people - the ones who\n    weren't. All those other people might be 1 million users who\n    share one fingerprint or 500,000 distinct fingerprints shared by\n    2 people each - that's quite the difference.\n\n    We use the\n    [Simpson\n    Index](https://en.wikipedia.org/wiki/Diversity_index) (aka the\n    [Herfindahl–Hirschman\n    Index](https://en.wikipedia.org/wiki/Herfindahl%E2%80%93Hirschman_index)) to quantify that difference and compare different\n    sets of cohorts (shared fingerprints). This lets us compare\n    how effective different fingerprinting defenses are based not\n    just on how much uniqueness they eliminate but how much they\n    collapse groups of users into each other.\n\n    For example, two different protections can each lower the share\n    of unique users by about 2.1 percentage points. But one improves\n    the cohort score by 56%, while another raises it by only 13%.\n    **Judged on uniqueness alone, two defenses can look equally\n    effective; while the cohort improvement varies significantly.**\n\nThe same protection performs very differently per platform\n:   Protections are not uniformly impactful, and the variation is\n    not small. As an example, clamping the available screen\n    resolution takes 25.3 percentage points off cumulative macOS\n    uniqueness but between 0 and 3 points off the other three\n    platforms. Averaged across all users, that 25-point drop\n    shrinks to 3.5 - **a protection that matters enormously to one\n    platform can look marginal in the aggregate**.\n\n    This is yet another reason why it's so important to segment and\n    analyze data on a per-operating-system basis.\n\nStandardization and randomization go hand-in-hand\n:   Making users report the same values makes them look more alike,\n    but it never makes everyone look the same. Some users will still\n    stand out, and that gets worse on websites with small visitor\n    populations. Standardization improves the situation, but it\n    should still be paired with randomization. Randomization is very\n    strong when it works, but can always be detected and bypassed.\n    **Neither can carry the burden of providing privacy on its\n    own.**\n\nThe data shows why values differ, not just that they do\n:   Correlating fingerprinting values with users' hardware and other\n    characteristics let us identify why certain measurements differed\n    between users. **Knowing the root cause lets us build\n    protections that reach much further, and protect small groups of\n    users who would otherwise stand out as a minority.**\n\n    Addressing these differences has already produced more consistent\n    outputs for\n    [image\n    rendering](https://bugzilla.mozilla.org/show_bug.cgi?id=2046791),\n    [Web\n    Audio](https://bugzilla.mozilla.org/show_bug.cgi?id=2036977), and\n    [mathematical\n    operations](https://bugzilla.mozilla.org/show_bug.cgi?id=2037094) on Android, with more root causes being investigated.\n\nUsers don't connect website breakage to the setting\n:   **Breakage from a strict mode is delayed, and the delay hides the cause.**\n    Turning on a stricter mode is a one-time act - but the broken website\n    arrives later and nothing connects the two. From the user's\n    perspective, this website doesn't work in Firefox and they have\n    no idea why and are frustrated. Report after report we see\n    people posting for help, confused - despite having taken\n    specific action to enable a hidden setting opting them into\n    strict fingerprinting protections.\n\nProtections that ship have to keep the web working\n:   Users don't realize that a setting they changed caused the breakage,\n    and don't know how to fix it - so a protection that breaks the web\n    breaks the browser along with it.\n    We selected and tuned mitigations on both measured impact\n    *and* compatibility. **For a default, compatibility\n    isn't a trade-off against privacy - it's a precondition for it.**\n\nThe hard part is finding what breaks\n:   The web is vast, and even the most innocuous change is likely to break\n    something somewhere. Some breakage can be anticipated, which helps\n    design, testing, and detection. **The other breakage gets fixed only if a\n    chain of luck holds: someone has to recognize the breakage, report it,\n    reproduce it on the right hardware, and get it to the right\n    person.**\n\nWe are deliberately not publishing everything we could\n:   Commercial fingerprinters evolve continuously, and they borrow\n    openly from the community that studies them. Sites built to\n    teach people about fingerprinting - CreepJS among them - get\n    picked up and used outright by fingerprinting products.\n\n    So we have been deliberately conservative. There are analyses\n    this dataset would support that we have not run, data we could\n    have released that we have not, and offensive angles on this\n    work that we have not pursued. That is a choice rather than an\n    oversight: **we would rather leave a question open than make it\n    easier to build a better fingerprinter**.\n\n## Test your own browser\n\nThis interactive test re-implements individual measurements as close\nas possible to how Firefox's own collection code performed them,\nruns in your browser, and compares each result against the published\naggregate data. For every characteristic you can see your value and how many\nof the ~25 million measured browsers share it - overall and within\nyour operating system.\n\n[Run the browser test\n→](https://fingerprinting.study/test/)\n\nThe test runs entirely in your browser. All comparisons happen\nlocally, against our static reference files.\nNo data is stored or submitted to us.\n\nIf the test, or anything else on this site, looks wrong, please\n[report it as an issue on GitHub](https://github.com/tomrittervg/fingerprinting_paper_artifacts/issues).\nFor a problem with the test, include the diagnostics from its\n**Copy diagnostics** button, or, if you would rather\nnot post your measured values publicly, email them to\n[tom@ritter.vg](mailto:tom@ritter.vg).\n\n## Paper\n\nThe extended version of the paper is available as a PDF.\n\n[Read the extended version\n(PDF) →](https://fingerprinting.study/fingerprinting-in-the-wild-extended.pdf)\n\n## Full appendix data\n\nThe paper's appendices contain a wealth of data locked behind a\nreasonable page limit and PDF formatting. This full data page\ncarries all of it - every rank for every operating system and\nall three modeled services, the complete protection chains,\nand the paper's figures regenerated at print resolution.\n\n[Open the full data\n→](https://fingerprinting.study/appendix.html)\n\n## Data collection\n\nOur measurements come from Firefox telemetry, specifically the\n`user-characteristics` ping, a dedicated telemetry ping.\nUnlike normal telemetry pings that are sent regularly, this ping\nmust be explicitly enabled (which has happened twice, once in 2025\nand once in 2026), and sends one data report before becoming dormant\nagain.\n\nThis ping carries no client ID, only a UUID unique to the ping used\nto de-duplicate users. It cannot be keyed to any other telemetry\nsubmissions. It is submitted over [Oblivious HTTP](https://www.rfc-editor.org/rfc/rfc9458)\n(OHTTP): a relay operated independently of Mozilla forwards the\nencrypted submission, so the relay can see the sender's IP address but\nnot the contents, and Mozilla receives the contents but not the IP\naddress. Access to the collected data was restricted to four authorized\nMozilla employees. The data was retained for 90 days and has since been deleted.\n\n[Read about the data collection\n→](https://docs.google.com/document/d/1m_j0BQEprQleRHZ7tVT7mG-krc8UA171GD5Vl6gZbL0/edit?tab=t.0)\n\n## Aggregate data & analysis code\n\nWe release four types of data:\n\n1. A list of fingerprints generated for different phases of protections, per modeled fingerprinter, with the user's OS and locale unmasked.\n2. Aggregate tables of the frequency of value per metric, split by OS\n3. Aggregate tables of the frequency of value per metric, correlated by other metrics' values, again split by OS\n4. A detailed analysis of rendered canvas images that allows comparison\n\nIn the fingerprint lists, the fingerprints cannot be recreated from\na website measuring user characteristics, due to the way the telemetry\nwas collected. After being generated, the files were shuffled to a new ordering,\nand any locale occurring fewer than 100 times was anonymized.\n\nIn the aggregate tables, any value observed fewer\nthan 100 times within a single per-OS population is withheld - so\nthe published data describes cohorts, never individuals.\n\nIn the canvas comparison, raw canvas renderings are provided, but they\nare devoid of any connection to other user data. Any unique canvas rendering\nis not provided; only the hash.\n\nFor verification and exploration, the code used to process the data and\nperform different analyses is available for review. It additionally contains\nthe specific fingerprinting scripts captured, reverse engineered, and\nmodeled for the paper.\n\n[Download the data\n→](https://fingerprinting.study/download.html)\n[Canvas explorer →](https://storage.googleapis.com/canvas-comparison/index.html)\n[Analysis code →](https://github.com/tomrittervg/fingerprinting_paper_artifacts)\n\n## Future Research\n\nFar more can be asked of this dataset than we have asked. Some of it\nwe chose not to pursue, some of it we simply did not think of, and\nsome of it is a better fit for someone who is not us.\n\nThere is a discussion board for proposing follow-up analyses. Two\nkinds are worth posting:\n\n- **Answerable from the published aggregate data.**\n  Anyone can run these - the per-operating-system\n  distributions are released, and the analysis toolkit is open. Post\n  the question anyway, so the work is not duplicated and so the\n  result lands somewhere others can find it.\n- **Needs the raw data.** These we would have to run\n  ourselves, since the underlying ping data is not something we can\n  ever release. Describe what you want to know and why! To be clear,\n  we *do not have the raw data anymore* but it is possible we\n  will perform a re-collection in the future which we could analyze.\n  We cannot promise to run everything, and some questions we will\n  decline because we do not wish to help build a better fingerprinter.\n\n[Open the discussion\nboard\n→](https://github.com/tomrittervg/fingerprinting_paper_artifacts/discussions)\n","body_html":"<h1 id=\"large-scale-measurement-and-real-world-mitigation-of-web-browser\">Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild</h1>\n<p>Tom Ritter1, Fatih Kilic2, Frederik\nBraun1, Elisa Luo3, Geoffrey M.\nVoelker3, Stefan Savage3, Christoph\nKerschbaumer1</p>\n<p>1Mozilla Corporation  ·  2University\nof Toronto  ·  3UC San Diego</p>\n<p>This is the companion site to a study Mozilla ran on browser\nfingerprinting. With the help of almost 25 million Firefox release\nusers, we measured how identifiable browsers really are, found which\ncharacteristics make them unique, and used that to design the\nfingerprinting protections that shipped in Firefox 143. Here you can\nread the paper and its key findings, browse the full appendix data\nthe PDF had no room for, download the privacy-preserving aggregate\ndata and analysis code, learn how the data was collected, and run an\ninteractive test that compares your own browser against the measured\npopulation. If you have a question the data could answer, there is\nalso a discussion board for proposing follow-up research.</p>\n<p>Browser fingerprinting is a known tracking mechanism on the modern\nweb, but its real-world effectiveness toward users has been hard\nto measure. To build robust protections, we measured uniqueness\nacross a large, unbiased population, to determined what individual\nmetrics cause users to be unique.\nThese measurements used multiple privacy preserving technologies,\nlike OHTTP, to ensure that\nno individual users can be identified or tracked across the web.\nThe protections we built into Firefox avoid web compatibility\nproblems while dropping the percentage of uniquely identifiable\nusers from 75% to 46%.</p>\n<p><a href=\"https://fingerprinting.study/test/\" rel=\"nofollow ugc noopener\">Test your own browser\n→</a>\n<a href=\"#paper\">Read the paper</a>\n<a href=\"https://fingerprinting.study/appendix.html\" rel=\"nofollow ugc noopener\">Full appendix data</a>\n<a href=\"https://fingerprinting.study/download.html\" rel=\"nofollow ugc noopener\">Download the data</a></p>\n<p><img src=\"https://fingerprinting.study/assets/figures/protection_chain.light.png\" alt=\"Line chart. The share of users with a unique fingerprint\n                falls as each of the protections Firefox 143 shipped is\n                added, for three modeled fingerprinters. Fingerprint Pro\n                falls from 75.7% to 46.8%, Signifyd from 60.9% to 38.5%,\n                and Akamai from 18.2% to 9.0%.\" loading=\"lazy\" decoding=\"async\" referrerpolicy=\"no-referrer\" /></p>\n<p>Shows the percentage of users seen as unique by three commercial\nfingerprinters. Each step adds one protection on top of all the ones\nbefore it, weighted across the four operating systems.</p>\n<h2 id=\"paper-abstract\">Paper abstract</h2>\n<p>Web browser fingerprinting has become a key tracking mechanism on\nthe modern web, yet its real-world effectiveness and which of the\nexisting mitigations offer value at scale remain insufficiently\nunderstood.</p>\n<p>In this paper, we present a large-scale study of browser\nfingerprinting based on real-world telemetry from almost 25 million\nFirefox release users. This dataset enables us to quantify user\nuniqueness at population scale and identify which characteristics\ncontribute most to this uniqueness.</p>\n<p>Our analysis shows fingerprintability is largely driven by a small\nset of high-impact characteristics and their combinations, rather\nthan the accumulation of many independent measurements. Building on\nthese insights, we propose and evaluate fingerprinting protections\nbased on their impact on uniqueness and cohort structure, and design\ndata-driven mitigations.</p>\n<p>These protections, shipped in Firefox 143, reduce the fraction of\nuniquely identifiable users from 75% to 46% while maintaining web\ncompatibility, demonstrating that practical reductions in\nfingerprintability are achievable in a mainstream browser.</p>\n<h2 id=\"our-key-findings\">Our key findings</h2>\n<p>Here&#39;s what we want to present to you with this paper and the\nbroader scope of our work.</p>\n<p>Uniqueness is caused by a small number of metrics - not one or thirty\n:   Fingerprinting test sites have a seemingly endless list of\n    measurements they can pull from your browser. If a user has a\n    unique fingerprint, it is typically not because of a single\n    measurement, but it is <em>also</em> not because of 20 or 30\n    measurements. <strong>It is a small number of metrics that make\n    users unique</strong>.</p>\n<pre><code>In fact, it&#39;s very close to the\n[80/20\nrule](https://en.wikipedia.org/wiki/Pareto_principle). Specifically, of all the users seen as unique by\nFingerprint Pro (the most discerning fingerprinter we model) -\n80% of them are unique after only 8 metrics - while the\nfingerprinter collects a total of over 50 metrics. Signifyd\ncollects more metrics, but identifies fewer users as unique:\n80% of the users it sees as unique will be identified after 9\nmetrics despite collecting over 80.</code></pre>\n<p>The numbers only mean anything per operating system\n:   Whenever someone cites a single percentage when describing &quot;How\n    many users are unique&quot; - you should be suspicious. Because\n    <strong>this collapses multiple different populations that\n    behave very differently</strong>.</p>\n<pre><code>Firstly, as a baseline, your operating system is always going\nto be exposed to a fingerprinter. We are always working within\nthe constraint of web compatibility, and lying about the\nuser&#39;s operating system breaks shortcut keys and gives\nyou the wrong downloads among other problems. So if\nwe *know* the website is going to learn your real OS\nanyway, we can use it to segment the population and analyze it\nseparately.\n\nAnd when you do you see that each operating system behaves\ndifferently because they have different important metrics.\nAdmittedly, it&#39;s very difficult to give a concise summary of\nwork and also provide four sets of numbers - so even we fall\nprey to the trap of providing a single number, as we did in\nthe prior paragraph. But *every time* we have a single\npercentage, somewhere we have that number broken down by\noperating system, and in doing so, you will see that each\nbehaves differently.</code></pre>\n<p>Entropy is not the measurement you want\n:   It&#39;s so, so easy to take all the results for a metric and just\n    calculate Shannon entropy - it&#39;s a standard yardstick in this\n    literature. But as we show - <strong>a high Shannon entropy does not\n    reliably predict fingerprinting impact</strong>.</p>\n<p>Cohort size, not just uniqueness\n:   Previous analyses calculated how many people were unique, but\n    they did not analyze the <em>other</em> people - the ones who\n    weren&#39;t. All those other people might be 1 million users who\n    share one fingerprint or 500,000 distinct fingerprints shared by\n    2 people each - that&#39;s quite the difference.</p>\n<pre><code>We use the\n[Simpson\nIndex](https://en.wikipedia.org/wiki/Diversity_index) (aka the\n[Herfindahl–Hirschman\nIndex](https://en.wikipedia.org/wiki/Herfindahl%E2%80%93Hirschman_index)) to quantify that difference and compare different\nsets of cohorts (shared fingerprints). This lets us compare\nhow effective different fingerprinting defenses are based not\njust on how much uniqueness they eliminate but how much they\ncollapse groups of users into each other.\n\nFor example, two different protections can each lower the share\nof unique users by about 2.1 percentage points. But one improves\nthe cohort score by 56%, while another raises it by only 13%.\n**Judged on uniqueness alone, two defenses can look equally\neffective; while the cohort improvement varies significantly.**</code></pre>\n<p>The same protection performs very differently per platform\n:   Protections are not uniformly impactful, and the variation is\n    not small. As an example, clamping the available screen\n    resolution takes 25.3 percentage points off cumulative macOS\n    uniqueness but between 0 and 3 points off the other three\n    platforms. Averaged across all users, that 25-point drop\n    shrinks to 3.5 - <strong>a protection that matters enormously to one\n    platform can look marginal in the aggregate</strong>.</p>\n<pre><code>This is yet another reason why it&#39;s so important to segment and\nanalyze data on a per-operating-system basis.</code></pre>\n<p>Standardization and randomization go hand-in-hand\n:   Making users report the same values makes them look more alike,\n    but it never makes everyone look the same. Some users will still\n    stand out, and that gets worse on websites with small visitor\n    populations. Standardization improves the situation, but it\n    should still be paired with randomization. Randomization is very\n    strong when it works, but can always be detected and bypassed.\n    <strong>Neither can carry the burden of providing privacy on its\n    own.</strong></p>\n<p>The data shows why values differ, not just that they do\n:   Correlating fingerprinting values with users&#39; hardware and other\n    characteristics let us identify why certain measurements differed\n    between users. <strong>Knowing the root cause lets us build\n    protections that reach much further, and protect small groups of\n    users who would otherwise stand out as a minority.</strong></p>\n<pre><code>Addressing these differences has already produced more consistent\noutputs for\n[image\nrendering](https://bugzilla.mozilla.org/show_bug.cgi?id=2046791),\n[Web\nAudio](https://bugzilla.mozilla.org/show_bug.cgi?id=2036977), and\n[mathematical\noperations](https://bugzilla.mozilla.org/show_bug.cgi?id=2037094) on Android, with more root causes being investigated.</code></pre>\n<p>Users don&#39;t connect website breakage to the setting\n:   <strong>Breakage from a strict mode is delayed, and the delay hides the cause.</strong>\n    Turning on a stricter mode is a one-time act - but the broken website\n    arrives later and nothing connects the two. From the user&#39;s\n    perspective, this website doesn&#39;t work in Firefox and they have\n    no idea why and are frustrated. Report after report we see\n    people posting for help, confused - despite having taken\n    specific action to enable a hidden setting opting them into\n    strict fingerprinting protections.</p>\n<p>Protections that ship have to keep the web working\n:   Users don&#39;t realize that a setting they changed caused the breakage,\n    and don&#39;t know how to fix it - so a protection that breaks the web\n    breaks the browser along with it.\n    We selected and tuned mitigations on both measured impact\n    <em>and</em> compatibility. <strong>For a default, compatibility\n    isn&#39;t a trade-off against privacy - it&#39;s a precondition for it.</strong></p>\n<p>The hard part is finding what breaks\n:   The web is vast, and even the most innocuous change is likely to break\n    something somewhere. Some breakage can be anticipated, which helps\n    design, testing, and detection. <strong>The other breakage gets fixed only if a\n    chain of luck holds: someone has to recognize the breakage, report it,\n    reproduce it on the right hardware, and get it to the right\n    person.</strong></p>\n<p>We are deliberately not publishing everything we could\n:   Commercial fingerprinters evolve continuously, and they borrow\n    openly from the community that studies them. Sites built to\n    teach people about fingerprinting - CreepJS among them - get\n    picked up and used outright by fingerprinting products.</p>\n<pre><code>So we have been deliberately conservative. There are analyses\nthis dataset would support that we have not run, data we could\nhave released that we have not, and offensive angles on this\nwork that we have not pursued. That is a choice rather than an\noversight: **we would rather leave a question open than make it\neasier to build a better fingerprinter**.</code></pre>\n<h2 id=\"test-your-own-browser\">Test your own browser</h2>\n<p>This interactive test re-implements individual measurements as close\nas possible to how Firefox&#39;s own collection code performed them,\nruns in your browser, and compares each result against the published\naggregate data. For every characteristic you can see your value and how many\nof the ~25 million measured browsers share it - overall and within\nyour operating system.</p>\n<p><a href=\"https://fingerprinting.study/test/\" rel=\"nofollow ugc noopener\">Run the browser test\n→</a></p>\n<p>The test runs entirely in your browser. All comparisons happen\nlocally, against our static reference files.\nNo data is stored or submitted to us.</p>\n<p>If the test, or anything else on this site, looks wrong, please\n<a href=\"https://github.com/tomrittervg/fingerprinting_paper_artifacts/issues\" rel=\"nofollow ugc noopener\">report it as an issue on GitHub</a>.\nFor a problem with the test, include the diagnostics from its\n<strong>Copy diagnostics</strong> button, or, if you would rather\nnot post your measured values publicly, email them to\n<a href=\"mailto:tom@ritter.vg\">tom@ritter.vg</a>.</p>\n<h2 id=\"paper\">Paper</h2>\n<p>The extended version of the paper is available as a PDF.</p>\n<p><a href=\"https://fingerprinting.study/fingerprinting-in-the-wild-extended.pdf\" rel=\"nofollow ugc noopener\">Read the extended version\n(PDF) →</a></p>\n<h2 id=\"full-appendix-data\">Full appendix data</h2>\n<p>The paper&#39;s appendices contain a wealth of data locked behind a\nreasonable page limit and PDF formatting. This full data page\ncarries all of it - every rank for every operating system and\nall three modeled services, the complete protection chains,\nand the paper&#39;s figures regenerated at print resolution.</p>\n<p><a href=\"https://fingerprinting.study/appendix.html\" rel=\"nofollow ugc noopener\">Open the full data\n→</a></p>\n<h2 id=\"data-collection\">Data collection</h2>\n<p>Our measurements come from Firefox telemetry, specifically the\n<code>user-characteristics</code> ping, a dedicated telemetry ping.\nUnlike normal telemetry pings that are sent regularly, this ping\nmust be explicitly enabled (which has happened twice, once in 2025\nand once in 2026), and sends one data report before becoming dormant\nagain.</p>\n<p>This ping carries no client ID, only a UUID unique to the ping used\nto de-duplicate users. It cannot be keyed to any other telemetry\nsubmissions. It is submitted over <a href=\"https://www.rfc-editor.org/rfc/rfc9458\" rel=\"nofollow ugc noopener\">Oblivious HTTP</a>\n(OHTTP): a relay operated independently of Mozilla forwards the\nencrypted submission, so the relay can see the sender&#39;s IP address but\nnot the contents, and Mozilla receives the contents but not the IP\naddress. Access to the collected data was restricted to four authorized\nMozilla employees. The data was retained for 90 days and has since been deleted.</p>\n<p><a href=\"https://docs.google.com/document/d/1m_j0BQEprQleRHZ7tVT7mG-krc8UA171GD5Vl6gZbL0/edit?tab=t.0\" rel=\"nofollow ugc noopener\">Read about the data collection\n→</a></p>\n<h2 id=\"aggregate-data-analysis-code\">Aggregate data &amp; analysis code</h2>\n<p>We release four types of data:</p>\n<ol><li>A list of fingerprints generated for different phases of protections, per modeled fingerprinter, with the user&#39;s OS and locale unmasked.</li><li>Aggregate tables of the frequency of value per metric, split by OS</li><li>Aggregate tables of the frequency of value per metric, correlated by other metrics&#39; values, again split by OS</li><li>A detailed analysis of rendered canvas images that allows comparison</li></ol>\n<p>In the fingerprint lists, the fingerprints cannot be recreated from\na website measuring user characteristics, due to the way the telemetry\nwas collected. After being generated, the files were shuffled to a new ordering,\nand any locale occurring fewer than 100 times was anonymized.</p>\n<p>In the aggregate tables, any value observed fewer\nthan 100 times within a single per-OS population is withheld - so\nthe published data describes cohorts, never individuals.</p>\n<p>In the canvas comparison, raw canvas renderings are provided, but they\nare devoid of any connection to other user data. Any unique canvas rendering\nis not provided; only the hash.</p>\n<p>For verification and exploration, the code used to process the data and\nperform different analyses is available for review. It additionally contains\nthe specific fingerprinting scripts captured, reverse engineered, and\nmodeled for the paper.</p>\n<p><a href=\"https://fingerprinting.study/download.html\" rel=\"nofollow ugc noopener\">Download the data\n→</a>\n<a href=\"https://storage.googleapis.com/canvas-comparison/index.html\" rel=\"nofollow ugc noopener\">Canvas explorer →</a>\n<a href=\"https://github.com/tomrittervg/fingerprinting_paper_artifacts\" rel=\"nofollow ugc noopener\">Analysis code →</a></p>\n<h2 id=\"future-research\">Future Research</h2>\n<p>Far more can be asked of this dataset than we have asked. Some of it\nwe chose not to pursue, some of it we simply did not think of, and\nsome of it is a better fit for someone who is not us.</p>\n<p>There is a discussion board for proposing follow-up analyses. Two\nkinds are worth posting:</p>\n<ul><li><p><strong>Answerable from the published aggregate data.</strong></p><p>Anyone can run these - the per-operating-system\ndistributions are released, and the analysis toolkit is open. Post\nthe question anyway, so the work is not duplicated and so the\nresult lands somewhere others can find it.</p></li><li><p><strong>Needs the raw data.</strong> These we would have to run</p><p>ourselves, since the underlying ping data is not something we can\never release. Describe what you want to know and why! To be clear,\nwe <em>do not have the raw data anymore</em> but it is possible we\nwill perform a re-collection in the future which we could analyze.\nWe cannot promise to run everything, and some questions we will\ndecline because we do not wish to help build a better fingerprinter.</p></li></ul>\n<p><a href=\"https://github.com/tomrittervg/fingerprinting_paper_artifacts/discussions\" rel=\"nofollow ugc noopener\">Open the discussion\nboard\n→</a></p>","headings":[{"level":1,"text":"Large-Scale Measurement and Real-World Mitigation of Web Browser Fingerprinting in the Wild","id":"large-scale-measurement-and-real-world-mitigation-of-web-browser"},{"level":2,"text":"Paper abstract","id":"paper-abstract"},{"level":2,"text":"Our key findings","id":"our-key-findings"},{"level":2,"text":"Test your own browser","id":"test-your-own-browser"},{"level":2,"text":"Paper","id":"paper"},{"level":2,"text":"Full appendix data","id":"full-appendix-data"},{"level":2,"text":"Data collection","id":"data-collection"},{"level":2,"text":"Aggregate data & analysis code","id":"aggregate-data-analysis-code"},{"level":2,"text":"Future Research","id":"future-research"}]}}