Same Browser Fingerprint Across Direct, HTTP, and SOCKS. Only the Document IPs Moved.

Browser fingerprinting research: stable browser surfaces beside changing document IP labels across direct and proxy paths

A fresh Chromium context does not mint a new browser identity—and neither does routing document fetches through a residential HTTP or SOCKS proxy. In our cloud tests, eighteen isolated Playwright Chromium visits across three matched network conditions produced one composite fingerprint hash, while independently fetched document IP echoes shifted from cloud datacenter labels (US / AS14618 Amazon.com, Inc.) on direct egress to residential labels (BR / AS28142…) on both ProxySeller HTTP and SOCKS paths.

We measured researcher-operated fingerprint surfaces—canvas, WebGL, audio, navigator, screen, timezone, font metrics, storage availability, media-device availability, and WebRTC candidate counts—inside headless Chromium 131 on a Cursor-hosted USA VM. Amendment A1 activated ProxySeller residential comparison legs after Runtime Secrets injection. CometVPN remains VPN-only context, not a measured tunnel.

18 / 18Matched successful visits
1Shared composite hash
3Egress conditions
12 / 12Scored surfaces stable

Playwright Chromium 131.0.6778.33, headless, locale en-US, timezone UTC, viewport 1360×900, fixture delivery page.setContent. Matched control 20261117T175322Z-cloud-direct; HTTP 20261117T174906Z-proxy-us-http; SOCKS 20261117T175129Z-proxy-us-socks.

Ce que nous avons mesuré

The frozen protocol lives at research/browser-fingerprinting-study/protocol.md (Amendment A1). Each visit loaded a researcher fixture that hashed canvas and WebGL output, scored an OfflineAudioContext metric, measured ten font strings, recorded navigator/screen/timezone/storage fields, gathered WebRTC ICE candidates with Cloudflare STUN only, and fetched four public document echoes with abort timeouts.

Residential proxies denied loopback HTTP (503 Target host denied), so every matched condition—including cloud-direct—uses Playwright page.setContent for fixture delivery. That keeps origin-bound fields identical across the matrix. SOCKS uses a local proxy-chain anonymizer because Chromium rejects SOCKS5 username/password.

Matched A1 collection windows (UTC, 17 November 2026): HTTP 17:49:16–17:50:13; SOCKS 17:51:39–17:52:33; matched cloud-direct 17:53:31–17:54:12. The protocol’s supplementary proxy-dead cell was not activated in A1.

For product-facing explanations of the same ideas, see ce que les empreintes digitales du navigateur est, types d'empreintes digitales du navigateur, HTTP vs SOCKS proxies, and our free canvas fingerprint tester.

Cloud-direct

6/6 visits · country US

AS14618 Amazon.com, Inc.

ProxySeller HTTP

6/6 visits · country BR

AS28142 DIGITAL DESIGN SERVIÇOS DE TELECOMUNICAÇÕES EIRELI

ProxySeller SOCKS

6/6 visits · country BR

AS28142 DIGITAL DESIGN SERVIÇOS DE TELECOMUNICAÇÕES EIRELI

Finding 1: eighteen matched visits, one composite fingerprint

All eighteen matched visits shared one composite hash across cloud-direct, ProxySeller HTTP and ProxySeller SOCKS. Closing a context—or changing the document-fetch path—did not rotate canvas, WebGL, audio, navigator, screen, font-metric, timezone or WebRTC count fields. That aligns with cookies versus local storage versus fingerprinting: storage isolation is not rendering isolation.

Shared composite hash: b661fd318fd710e98a86095230b3c5d423829410cb55806e1a38ecad1a1dd1f1

Canvas hash (unchanged from the historical localhost run): 57e77ee7b491b06625fd9d58be3ddcdc3a61ac9df3fafadd4ef1ab9e3fcc59f5

Surface stability across matched visits

Each bar is unique values observed within a condition. Full width means one unique value (stable). The same surfaces also matched across direct, HTTP and SOCKS.

compositeHash1 unique / 6
canvas1 unique / 6
webgl.dataHash1 unique / 6
webgl.renderer1 unique / 6
audio.absSum1 unique / 6
timezone1 unique / 6
navigator.userAgent1 unique / 6
navigator.platform1 unique / 6
navigator.hardwareConcurrency1 unique / 6
screen1 unique / 6
fonts1 unique / 6
webrtc.counts1 unique / 6
Twelve scored surfaces were stable within and across conditions. This is repeatability on one stack, not a uniqueness percentile.
Fixture screenshot after cloud-direct visit p1-v1 showing composite hash summary
cloud-direct · visit p1-v1
Fixture screenshot after ProxySeller HTTP visit p1-v1 showing the same composite hash summary
proxy-us-http · visit p1-v1
Fixture screenshot after ProxySeller SOCKS visit p1-v1 showing the same composite hash summary
proxy-us-socks · visit p1-v1

Fixture screenshots show only hash / renderer / timezone summaries—no raw IPs.

Finding 2: document IP labels moved with the proxy; the fingerprint did not

On matched cloud-direct, all four echo kinds succeeded (24 / 24) but returned 7 distinct salted labels total while staying in US / AS14618 Amazon.com, Inc.. Per-kind unique labels: ipify 4, api64 4, ipinfo 5, Cloudflare 4. Within-visit agreement was incomplete (ipify↔Cloudflare 0/6; ipify↔ipinfo 1/6). On both ProxySeller legs in this session, echoes moved to BR / AS28142 DIGITAL DESIGN SERVIÇOS DE TELECOMUNICAÇÕES EIRELI, with one salted label across kinds and full within-visit ipify agreement. The composite fingerprint hash did not move.

That is the measured core of pourquoi changer votre IP ne change pas l'empreinte de votre navigateur et fait une VPN changer votre empreinte de navigateur?—now shown with residential HTTP and SOCKS comparison legs, not only within-datacenter label churn. Echo geo/org labels are session observations, not a ProxySeller inventory claim. Within-datacenter label churn on direct egress is itself evidence that echo endpoints are not a single ground-truth address oracle.

ÉtatEcho successesUnique salted labelsCountry / org
Cloud-direct24 / 247US · AS14618 Amazon.com, Inc.
ProxySeller HTTP24 / 241BR · AS28142 DIGITAL DESIGN SERVIÇOS DE TELECOMUNICAÇÕES EIRELI
ProxySeller SOCKS24 / 241BR · AS28142 DIGITAL DESIGN SERVIÇOS DE TELECOMUNICAÇÕES EIRELI

Cloud-direct: echo labels

ipify vs Cloudflare0 / 6
ipify vs ipinfo1 / 6
Salted labels only. Cloud-direct country US; composite unchanged.

ProxySeller HTTP: echo labels

ipify vs Cloudflare6 / 6
ipify vs ipinfo6 / 6
Salted labels only. ProxySeller HTTP country BR; composite unchanged.

ProxySeller SOCKS: echo labels

ipify vs Cloudflare6 / 6
ipify vs ipinfo6 / 6
Salted labels only. ProxySeller SOCKS country BR; composite unchanged.

Finding 3: automation and software-GPU signals were part of the stable identity

The stable navigator subset included webdriver: true, pluginsLength: 0, and a HeadlessChrome user agent. WebGL reported SwiftShader. Hardware concurrency was 4. Under setContent delivery, deviceMemory was absent and storage APIs reported unavailable—identical on direct and proxied visits. Accuracy context: la précision de l'empreinte du navigateur.

MarkerObserved value (all matched conditions)
webdrivertrue
pluginsLength0
WebGL rendererANGLE (Google, Vulkan 1.3.0 (SwiftShader Device (Subzero) (0x0000C0DE)), SwiftShader driver)
audio.absSum33.6482334
deviceMemorynull
localStorage / sessionStoragefalse / false
media.availablefalse
canvas hash57e77ee7b491b06625fd9d58be3ddcdc3a61ac9df3fafadd4ef1ab9e3fcc59f5

Finding 4: WebRTC candidate structure repeated

Each matched visit returned two ICE candidates—one host and one srflx—after the eight-second gather timeout, with no relay candidates. Address-class labels stayed one none and one public-or-other-v4. Raw candidate IPs are not published. Structure repeated across direct, HTTP and SOCKS; UDP STUN through a browser proxy was not separately instrumented beyond those counts. See also ce qui est une fuite WebRTC et IP fuites, WebRTC et IPv6.

Finding 5: HTTP and SOCKS agreed with each other

ProxySeller HTTP (native Playwright auth) and ProxySeller SOCKS (local anonymizer → upstream SOCKS5) produced the same composite hash, the same rendering surfaces, and the same echo country/org labels in this session. Design JSON records proxyViaLocalAnonymizer: true on the SOCKS leg. Related: HTTP vs SOCKS proxies et types de serveurs proxy.

What ProxySeller and CometVPN mean in this package

ProxySeller. Amendment A1 executed residential USA-protocol HTTP/SOCKS Chromium legs using injected Runtime Secret component names. Measurements are researcher fixtures and document echoes—not competitive uptime benchmarking. Evaluate ProxySeller as a product separately from this fingerprint matrix. Related: ProxySeller review.

CometVPN. CometVPN is VPN-only. This study does not use it as HTTP/SOCKS and does not claim tunnel measurements. Product context: CometVPN, our Revue CometVPNet ce que les VPN ne protègent pas.

A better way to check fingerprint assumptions

  1. Fix browser, viewport, locale, timezone and fixture delivery origin.
  2. Collect the same surface set in fresh contexts.
  3. Hash surfaces you can explain—not only a third-party mystery score.
  4. Record network identity separately, with redaction.
  5. Change one network condition only after the baseline is stable under the same delivery.
  6. If loopback fixtures fail through a residential proxy, change delivery and re-baseline—do not silently skip.

If you need intentional profile isolation rather than a tunnel, start from ce qu'est un navigateur anti-détecteur, comment vérifier votre empreinte digitale du navigateuret peut empêcher l'empreinte du navigateur.

Methods appendix: guards and failure modes

  • Env discovery records secret names only (evidence/env-discovery.json). A1 recorded all five component names as PRESENT; /dev/net/tun existed and was not treated as a CometVPN tunnel proof.
  • Credentials never enter public visit JSON, design JSON, screenshots or findings.
  • Echo fetches use 12s abort timeouts; media/audio/WebRTC stages have outer timeouts.
  • Failed pre-fix HTTP attempts (loopback denial, smoke timeout, missing SHA-256 fallback) are retained as excluded evidence.
  • Historical localhost cloud-direct composite dd8a9bcc6d71907feaad2e4233563690ebc655ccf0c76c73a03a651927f7ecc9 differs only on origin-bound fields and is excluded from matched totals.
  • Reproducibility: install Playwright Chromium from the study package.json, freeze protocol.md, run collect.mjs per condition with injected PROXYSELLER_* secrets for proxy legs, then analyze.py. Keep page.setContent delivery for matched comparisons.

Ce que cette étude ne peut vous dire

  • One automation stack. Headless Chromium 131 with SwiftShader on one USA Cursor VM is not desktop Chrome entropy.
  • One session window. Matched A1 legs ran within about five UTC minutes on 17 November 2026—no day-to-day drift measurement.
  • proxy-dead not run. Failure-mode proxy behavior remains unmeasured.
  • No CometVPN tunnel. Affiliate/product context only; CometVPN is VPN-only.
  • No population uniqueness et no ProxySeller ranking from echo geo/org labels.
  • WebRTC is structural only—candidate counts/classes, not a full proxy-egress leak verdict.
  • Opaque-origin fixture disables some storage/deviceMemory/media fields by design for matched delivery.
  • No provider logging claim, headed consumer browser, anti-detect profile, or commercial uniqueness-score primary sample.

Preuves, exclusions et divulgation

The public ZIP includes amended protocol, collector, fixture, analyzer, env-discovery (names only), matched-control and proxy-leg visits, design files, source snapshots, fixture screenshots, manifests, analysis and the markdown article. Raw IP bodies and salts remain private. No competitive ProxySeller uptime benchmark was run. proxy-dead was not activated.

Télécharger les preuves et reproduire le résultat

Amended protocol, collector, fixture, redacted visits for all three matched conditions, design files, source snapshots, fixture screenshots and SHA-256 manifests are included. Raw IPs stay private.

Foire aux questions

Sources et notes de recherche

Informations sur les sociétés affiliées : This article contains ProxySeller and CometVPN affiliate links. FreePolicyTools may earn a commission if you purchase through them. These offers are separate from the research findings and do not constitute a comparative product test or change the study’s methods, results or limitations.

Traduction automatique