Skip to content
iplogs.comSign up
Guide · Last updated 2026-11-19

AI and ML in VPN detection: what works, what's hype

Almost every IP-intelligence vendor now calls itself “AI-powered”. A few of them mean real machine learning. Most are describing signature matching, ASN classification and weights an engineer picked by hand, which are reasonable techniques with a new label on them. Below I sort the claims into categories, list the research I think is worth reading, and explain why we built a deterministic weighted pipeline instead of training a model.

By Mehmood Ali ·

Does AI or machine learning improve VPN detection?

Mostly not at the IP level, and the label is often applied to older techniques. Four things get called AI in IP intelligence: signature matching of JA3 and JA4 TLS fingerprints, hand-tuned weighted scoring, statistical heuristics such as the TCP-versus-TLS round-trip comparison from the SNITCH paper, and trained classifiers. Only trained classifiers are real machine learning. They help where inputs are too noisy for hand-written weights, such as behavioral velocity scoring, behavioral residential-proxy detection and traffic anomaly detection at CDN scale. Their cost is opacity: a single score with no per-signal provenance cannot explain why a customer was blocked. IPLogs uses deterministic weighted signals instead. The same input always gives the same verdict, every response lists the signals that fired with their weights, and the false-positive rate is published: 0.4% of 1,043 addresses on a public test. The trade-off is giving up interactions that a trained model might find.

What people actually mean by “AI” in this space

In practice the label covers four quite different things:

  • Signature matching. Comparing JA3/JA4 hashes against a database of known ones. The hash is computed from fixed fields in the TLS ClientHello, so there is no model to train. It's useful, and the underlying technique is decades old.
  • Hand-tuned weighted scoring. Signals such as a provider-list match, the ASN class, an active probe or a threat-intel hit, combined using weights an engineer chose. This is where we sit. The same input always gives the same verdict, and you can audit how it got there.
  • Statistical heuristics. The SNITCH paper's TCP-vs-TLS RTT differential is a good example: a mathematical model whose thresholds come from peer-reviewed research rather than from training on a labelled corpus. Where “heuristic” ends and “ML” begins is fuzzy. A workable test is whether the model can be fully explained in a paper. If it can, it isn't a black box.
  • Trained classifiers. A model that learned its weights from a labelled corpus, usually a gradient-boosted tree or a small neural net over behavioural features. Only this category really earns the “ML” label. A few vendors do it well, but many publish no error rate against any independent benchmark, so from the outside it is hard to tell which is which.

Where ML actually helps

A trained classifier is worth its opacity when the link between inputs and answer is too noisy for hand-written weights. That holds in three places I know of:

  • Behavioural velocity scoring. Request rate per account, geographical jumps within a session and the ratio of failed to successful logins interact non-linearly, and the right thresholds differ from product to product. Given a labelled corpus, a gradient-boosted tree usually beats a hand-tuned rule set here.
  • Behavioural residential-proxy detection. At the IP layer, a residential proxy and a real consumer look identical. Separating them takes session-level patterns: one account jumping between countries at the same time, high IP velocity, a session fingerprint that doesn't match. This is genuine ML territory, and some vendors are doing serious work on it.
  • Anomaly detection on traffic patterns. CDN-scale providers train models on a baseline of normal traffic and flag access patterns that deviate from it. It works for them because of the volume they see. With much less data it is hard to get right.

Where ML hurts: the black-box-score problem

Say your fraud team blocks a customer because “the model said so.” If the model returns one score and no per-signal provenance, nobody can tell whether the cause was a CGNAT collision, an Apple Private Relay miss, a residential proxy on a neighbour's router, or actual fraud. The customer has nothing specific to dispute, and whoever picks up the ticket has nothing to triage with. Opaque IP-reputation scores cause exactly the same trouble.

Why IPLogs uses deterministic weighted signals

Our pipeline combines many sources with hand-tuned weights. This is how it compares with the two ML approaches:

ApproachAuditable per verdictDrift over timePublished error rate
Black-box ML scoreNoYes, hard to detectUsually no
Transparent ML with feature importancePartialYes, slowerSometimes
Deterministic weighted signals (IPLogs)Yes, per-signal provenanceOnly when weights change, which is a deliberate PRYes, see /accuracy

We do give something up. A trained classifier can pick up interactions that hand-written weights miss, especially once session-level behavioural signals are added. In return, an IP-level detector with published weights and a published false-positive rate can be audited, and a classifier that can't explain individual decisions can't. Which matters more depends on the product. We went with auditability because the people calling the API are engineers too, and they have to defend each decision to their own users.

Cited research worth reading directly

  • SNITCH (NDSS 2025): the TCP vs TLS RTT differential. Background for the timing check we run. Check any specific figure against the published paper before you rely on it.
  • JA3 and JA4 fingerprint specifications. JA3 was first published by Salesforce engineering. JA4 is FoxIO's newer format, with a richer hash. Read the specs themselves rather than vendor summaries of them.
  • Anti-censorship protocol papers. Our blog covers REALITY, Xray, and AmneziaWG at the implementation level, but if you're building detection for them, start from the protocol papers.

FAQ

Does IPLogs use a machine-learning model?

No. The pipeline is deterministic. Each signal (provider list match, ASN classification, active protocol probe, RTT, JA3/JA4 fingerprint match) carries a documented weight, and the verdict is the weighted sum mapped to a bucket. No trained classifier is involved. We chose that on purpose, so that every verdict can be traced to the specific signals behind it.

Are JA3 and JA4 fingerprints 'AI'?

No. JA3 and JA4 are deterministic hashes of fields in the TLS ClientHello, and matching one against a database of known hashes is signature matching, the same family of technique antivirus software has used for decades. It is a useful technique. Calling it AI is marketing.

When is ML actually useful in IP intelligence?

When the relationship between the input features and the answer is too noisy for hand-written weights. Behavioural-velocity scoring is the usual example: it combines per-IP request rates, geographical jumps per user and account-takeover signals over time. A few vendors do this well, while others ship an unexplained score and call it AI.

Why does transparency matter for AI in detection?

A flagged user has no way to dispute a black-box score. If your fraud team blocks a customer because 'the model said so', nobody can tell whether the cause was a CGNAT collision, an Apple Private Relay mistake or a real fraud signal. When the verdict lists the signals that produced it, the team can work out which one it was.

What public research is worth reading on detection accuracy?

Start with the SNITCH paper from NDSS 2025. Section 5 covers the TCP-vs-TLS RTT differential used to detect middle-box relays; check any specific claim against the published paper before relying on it. JA3 and JA4 each have a published specification that is worth reading directly. Vendor blog posts about 'AI' are mostly marketing, so treat published research as the reliable source.

See the accuracy benchmark for our published false-positive rate, the JA3/JA4 glossary entry for the fingerprint primer, and the related blog post JA3 / JA4 TLS fingerprinting explained.