Skip to content
iplogs.comSign up
Published benchmark · Last updated

How accurate is IPLogs VPN detection?

On its most recent ground-truth benchmark, IPLogs records a 0.4% false-positive rate and no missed samples across a 1,043-IP labeled corpus, run against production. It detected 100% of published Mullvad relays and 100% of Tor exit nodes in that corpus.

Benchmark results

0.4%
False-positive rate
Clean residential / trusted IPs wrongly flagged as VPN.
0%
False-negative rate
None of the 714 VPN, Tor and datacenter samples missed. The VPN samples are listed servers.
100%
Mullvad relays detected
All 500 Mullvad WireGuard / OpenVPN relays in the corpus.
100%
Tor exits detected
All 54 running Tor exit nodes in the corpus.

What the errors were, in the 30 October 2026 run: of the 4 false positives, 3 are home or business lines that answered a VPN handshake (IKEv2 or OpenVPN), so a VPN server is running there. They now read suspicious rather than VPN: a router's own VPN server says nothing about whether the person using that line is on a VPN, but this corpus counts any non-clean verdict as an error. The fourth is an address the corpus files as plain CDN that published WARP egress ranges name as a WARP exit. The 29 October run had a fifth: a Cloudflare address read as WARP from a single upstream proxy flag, which no longer counts on Cloudflare's website addresses. Nothing was missed, but the VPN samples are servers that appear on published lists, so this run does not measure VPN servers no list names. The figures are reported as measured, not corrected.

A harder benchmark: 2,831 addresses

A second corpus, run against production on 2 November 2026, adds the places false positives usually hide: corporate VPN gateways, universities, mobile carriers and cloud servers. Its VPN side is 1,722 servers from 16 commercial providers. Most of those providers now publish lists iplogs.com reads, so that row mostly measures list coverage.

AddressesCountFlagged as VPN
Commercial VPN servers1,722100%
Home broadband3910
Mobile carriers1120
Universities910
Corporate gateways770
Cloud servers2002.0% (4)
CDN and public DNS800

Overall 0.4% of the 951 non-VPN addresses were flagged, all four of them cloud servers: three answered a live VPN or proxy handshake (IKEv2 or SOCKS5), and the fourth is tagged as a VPN server by open-port scans — a VPN or proxy really is running there, and the corpus files them as plain datacenter. Every home, mobile, university, corporate, CDN and public-DNS address read clean: a home line that answers a VPN handshake (a router's own VPN server) reads suspicious rather than VPN. Volunteer-run VPN servers are left out of the VPN figure: they are home connections that join and leave within hours, so a days-old sample mostly measures how fast a list goes stale. Checked against a live sample of those volunteer servers instead, 94 of 97 read as VPN (2 November).

iplogs.com result for 8.8.8.8: Google public DNS, clean, score 0 of 100, allowlisted reputation, with key findings showing no VPN, proxy or Tor evidence on a public DNS resolver network.
A hosting address that stays clean: 8.8.8.8 is recognised as public DNS infrastructure and scores 0/100.

On 2 November the same corpus in fast mode, which answers from memory in microseconds with no live probes, caught 99.7% of the commercial VPN servers (1,716 of 1,722) and flagged four cloud servers, three of them the same and the fourth a server that relays every connection to Cloudflare: once a full check has seen an address answer a VPN handshake, fast mode remembers it. The servers only a full check found were answering a VPN handshake while no longer listed by their provider.

Popular websites: 0.20% flagged

Range lists and open-proxy lists are where false positives start, so we also check the web and mail servers of the most-visited websites, which sit on the same hosting networks as VPN servers. On 2 November 2026, 10 of 4,999 such addresses read as VPN (fast mode, no live probes).

Four are web servers inside address ranges shared with VPN providers: real false positives. The other six do run VPN or proxy infrastructure under a popular name: four are listed as live proxy endpoints on open-proxy feeds, one is a VPN provider's own website on its server network, and one is a VPN service's rotating server domain. None was flagged by a VPN provider's own server list. The same check read 0.76% on the morning of 29 October, 0.22% that evening and 0.08% on 30 October. Two changes brought it down: listings found only on long-window proxy lists were weighted down, and neither those nor abuse-list hits can turn a hosting address into a VPN on their own.

What the corpus contains

The benchmark corpus is 1,043 IP addresses spanning five categories, chosen so that both the “should flag” and “should stay clean” sides are well represented:

  • Mullvad relays (500) — WireGuard and OpenVPN exit addresses of a large commercial VPN, in service at the time of the run. Ground truth for “is a VPN”.
  • Tor exit nodes (54) — exits running at the time of the run, IPv4 and IPv6.
  • Residential ISP addresses (282) — sampled per ISP. Ground truth for “should stay clean”; these are where false positives hurt most.
  • Datacenter / hosting IPs (160) — sampled per provider. Flagged as datacenter, not necessarily VPN, so the detector must distinguish the two.
  • CDN edge IPs (40) and public DNS resolvers (7) — sampled per provider (Cloudflare, Fastly, Akamai, Bunny). Must stay clean despite living in datacenter ASNs.

How the benchmark is run

  1. 1.Collect labeled IPs from authoritative sources — the servers the VPN and the Tor network list for themselves, and per-ISP / per-provider sampling for the clean and datacenter categories.
  2. 2.Validate every sample's ASN against live IP-to-ASN routing data. A residential, datacenter or CDN sample whose live ASN doesn't match its network is dropped, so a re-allocated address can't pose as an error.
  3. 3.Run against production — every corpus IP is checked against the live /v1/check API, not a lab build, so the numbers reflect what real callers get.
  4. 4.Compare to ground truth — each verdict is scored against the validated label to compute the false-positive, false-negative, and per-category true-positive rates above.
How a check worksFour kinds of evidence are checked in order and feed one score. 1: published server and relay lists (VPN servers, Tor exits, privacy relays, open proxies). 2: network and routing data (hosting and cloud networks, the ASN, routing and abuse reports). 3: live protocol handshakes such as OpenVPN, WireGuard, IKEv2 and SOCKS5, in full checks only. 4: the visitor's browser (time zone, language, WebRTC and connection timing), in self-checks only. Result: a score from 0 to 100 and a verdict of clean, suspicious, vpn_likely or vpn_detected.1Published server and relay listsVPN servers, Tor exits, privacy relays, open proxies2Network and routing dataHosting networks, ASN, routing, abuse reports3Live protocol handshakesOpenVPN, WireGuard, IKEv2, SOCKS5 (full checks)4The visitor's browserTime zone, language, WebRTC, timing (self-checks)Score and verdict0 to 100: clean, suspicious, vpn_likely, vpn_detected
How a check works: each layer adds evidence, and what matches adds up to one score. Fast and bulk checks skip layer 3; layer 4 runs only when visitors check their own connection.

Why the false-positive rate is the number that matters

For most teams, a false positive is far more expensive than a false negative: it blocks a paying customer, fails a real signup, or declines a legitimate payment. A detector that flags everything scores a perfect true-positive rate while making itself useless. The honest measure of a VPN detector is how rarely it flags a clean residential user — which is exactly why IPLogs publishes its 0.4% false-positive rate rather than only a true-positive headline.

Accuracy FAQ

How accurate is IPLogs VPN detection?

On its most recent ground-truth benchmark (30 October 2026, run against production), IPLogs records a 0.4% false-positive rate and no missed samples across a 1,043-IP labeled corpus. It detected 100% of the 500 published Mullvad relays and 100% of the 54 Tor exit nodes in that corpus.

What is a false positive in VPN detection?

A false positive is a clean residential or trusted IP that a detector wrongly flags as a VPN or proxy. False positives are the costly errors — they block real customers. IPLogs measures this directly and reports 0.4% on its ground-truth corpus.

What is a false negative in VPN detection?

A false negative is a real VPN, proxy, or Tor exit that a detector misses and marks as clean. In the latest run IPLogs missed none of the 714 VPN, Tor and datacenter samples. The VPN side of this corpus is Mullvad relays and Tor exits, servers that appear on published lists, so it measures list coverage; VPN servers that no list names are harder and are not measured here.

How is the benchmark corpus validated?

Every residential, datacenter and CDN sample has its origin ASN cross-checked against live IP-to-ASN routing data before it is labeled, and is dropped if the address now belongs to another network. Until 29 October 2026 only residential samples were checked, and every miss in those runs was a datacenter sample that had moved to another network.

Why don't other VPN detection services publish a false-positive rate?

Most commercial IP-intelligence vendors publish a confidence score but no measured error rate against a labeled corpus. Publishing a false-positive rate is falsifiable and invites scrutiny — IPLogs reports it because the figure is the number that decides whether you block real customers.

Reproduce or build on this

The detection methodology is documented openly and the aggregated VPN-provider dataset is published under CC-BY 4.0. See the API docs for the verdict schema, the datasets catalog for the downloadable feeds, and the comparison page for five control IPs you can use to test any VPN-detection API.