GitHub - adulau/ptrclassify: ptrclassify is a small, dependency-free Python library and CLI that infers likely IP usage and location from reverse-DNS PTR hostnames

GitHub

6 min read Original article ↗

ptrclassify is a small, dependency-free Python library and CLI that infers likely IP usage from reverse-DNS PTR hostnames.

It is intentionally heuristic and multi-label. PTR naming is operator-controlled and is not authoritative evidence of how an address is actually used. The output therefore includes a confidence score, the text that matched, and the rule IDs that produced each label.

Examples of orthogonal labels:

  • ptrclassify:allocation="dynamic", ptrclassify:allocation="static", ptrclassify:allocation="reserved"
  • ptrclassify:access="residential", ptrclassify:access="business", ptrclassify:access="mobile", ...
  • ptrclassify:translation="nat", ptrclassify:translation="cgnat"
  • ptrclassify:role="customer", ptrclassify:role="router", ptrclassify:role="broadband-aggregation", ...
  • ptrclassify:hosting="cloud", ptrclassify:hosting="datacenter", ptrclassify:hosting="hosting", ptrclassify:hosting="cdn"
  • ptrclassify:network="anycast", ptrclassify:network="dedicated-internet"
  • ptrclassify:organization="education", ptrclassify:organization="government", ptrclassify:organization="military"
  • ptrclassify:provider="amazon-aws", ptrclassify:provider="google-cloud", ...
  • ptrclassify:naming="ip-encoded", ptrclassify:naming="generic-reverse"

Install

Or install the published package from PyPI:

python -m pip install ptrclassify

For development:

python -m pip install -e .

Hyperscan is available as an optional high-throughput matching engine:

python -m pip install '.[hyperscan]'
ptrclassify --engine hyperscan 'ec2-3-151-166-120.us-east-2.compute.amazonaws.com.'

The default re engine has no third-party dependencies. The Hyperscan engine compiles all rule expressions into a single database and returns the same labels and evidence as the default engine.

API service

Install the API dependencies and start the FastAPI service:

python -m pip install '.[api]'
ptrclassify-api

The interactive Swagger UI is available at http://localhost:8000/docs, and FastAPI serves the OpenAPI document at http://localhost:8000/openapi.json. The same document is checked into this repository as openapi.json.

Classify one or more hostnames or complete PTR record lines in one request:

curl -X POST http://localhost:8000/lookup \
  -H 'content-type: application/json' \
  -d '{"records":["ec2-3-151-166-120.us-east-2.compute.amazonaws.com.","188.147.228.101.nat.umts.dynamic.t-mobile.pl."]}'

Library API

from ptrclassify import classify, PTRClassifier

labels = classify("188.147.228.101.nat.umts.dynamic.t-mobile.pl.")
for label in labels:
    print(label.value, label.confidence)

classifier = PTRClassifier()  # engine="re" (the default) or engine="hyperscan"
result = classifier.classify(
    "101.228.147.188.in-addr.arpa. PTR 188.147.228.101.nat.umts.dynamic.t-mobile.pl."
)
print(result.values())
print(result.to_dict())
for location in result.locations:
    print(location.city, location.region, location.country, location.confidence)

Expected high-confidence labels include:

ptrclassify:allocation="dynamic"
ptrclassify:access="mobile"
ptrclassify:translation="nat"

CLI

Single hostname:

ptrclassify 'ec2-3-151-166-120.us-east-2.compute.amazonaws.com.'

Complete DNS record:

ptrclassify '120.166.151.3.in-addr.arpa. PTR ec2-3-151-166-120.us-east-2.compute.amazonaws.com.'

File / JSONL:

ptrclassify --file tests/data/sample.ptr --json > classifications.jsonl

Just labels:

ptrclassify --values-only '86-45-50-202-dynamic.agg1.cab.bdt-fng.eircom.net.'

CSV enrichment

To classify every value in a CSV column named ptr, use the CSV enrichment script. It preserves the existing columns and appends a ptrclassify column containing the complete classification as compact JSON. It also appends a ptrclassify_labels column containing a compact JSON array of MISP machine-tag label values, which makes the most commonly used result available without parsing the full classification:

ptrclassify-csv input.csv output.csv
# Or, without installing the command:
python classify_ptr_csv.py input.csv output.csv

Use - (or omit both paths) to read from standard input or write to standard output. Alternate column names can be selected with --ptr-field, --labels-field, and --output-field:

ptrclassify-csv --ptr-field hostname --labels-field labels \
  --output-field classification input.csv output.csv

Rule model

Built-in rules live in ptrclassify/data/rules.json. Rules are regular-expression based and have this shape:

{
  "id": "access.mobile",
  "category": "access",
  "label": "mobile",
  "confidence": 0.98,
  "patterns": ["(?:^|[._-])mobile(?:[._-]|$)", "(?:^|[._-])umts(?:[._-]|$)"]
}

You can add private/local rules without modifying the package:

from ptrclassify import PTRClassifier

classifier = PTRClassifier(extra_rules=[{
    "id": "myisp.cgn",
    "category": "translation",
    "label": "cgnat",
    "confidence": 0.99,
    "patterns": [r"\\.cgn\\.example\\.net$"],
}])

Potential location extraction

The classifier can also decode a conservative set of operator-specific geographic tokens. Candidates are separate from taxonomy labels because a place is open-ended data rather than a classification category:

result = PTRClassifier().classify("pool-74-96-220-47.washdc.fios.verizon.net")
print(result.locations[0].to_dict())
# {'code': 'washdc', 'city': 'Washington', ..., 'confidence': 0.9}

Each candidate includes the original code, matched evidence, confidence and rule ID as well as any decoded city, region and ISO 3166-1 alpha-2 country. Hostnames ending in an ISO country-code TLD (for example, .pl or .de) also produce a lower-confidence country candidate. A country-code TLD reflects the DNS namespace and does not guarantee that the named host is physically in that country; generic TLDs such as .com and .net do not produce this hint. Built-in templates live in ptrclassify/data/location_rules.json. They are deliberately scoped to an operator suffix: a token such as ord56 on an unrelated domain is not treated as Chicago. Private conventions can be added with extra_location_rules:

The built-in infrastructure templates include AWS EC2 region names, Azure cloudapp.azure.com regions, and a curated set of Amazon CloudFront POP codes. These complement the access-network templates and make the location candidate useful for CDN and datacenter PTRs without treating generic IATA-like labels as locations.

classifier = PTRClassifier(extra_location_rules=[{
    "id": "location.example.site",
    "confidence": 0.95,
    "pattern": r"\.(?P<code>hq)\.example\.net$",
    "locations": {"hq": {"city": "Example City", "country": "ZZ"}},
}])

This follows the explainable, domain-specific rule strategy used by hostname geolocation systems: extract tokens in the context of a network's naming convention, then resolve them through a curated code dictionary. Backbone router templates for Hurricane Electric and Cogent apply the same approach to operator-scoped POP codes (for example, lax2 and lax01). A candidate describes where the operator says the named device or service belongs; it must not be interpreted as the subscriber's exact location or as a measurement.

Benchmark

After installing the optional dependency, compare steady-state lookup time on the bundled sample data with:

python benchmarks/lookup.py

Use --iterations, --repeat, or --file to change the workload. Classifier construction is deliberately excluded from the timed section so the result measures lookup throughput rather than one-time expression compilation.

Design notes

The design follows the same general idea used by Internet-topology work such as CAIDA Hoiho: operator naming conventions can be mined as evidence, but should be treated as inference rather than truth. This package focuses on usage/allocation/service classes instead of primarily extracting router geolocation.

Provider-specific rules are intentionally separated from generic tokens. This avoids dangerous inferences such as classifying every softbank... hostname as mobile simply because SoftBank also operates mobile networks.

For production enrichment, PTR classification is best combined with ASN/RDAP, BGP prefix data, geofeeds, known cloud/hosting prefixes, forward-confirmed reverse DNS (FCrDNS), and active/service observations.

References / prior art

  • RFC 8501, Reverse DNS in IPv6 for Internet Service Providers: discusses static, dynamic and dynamically generated reverse names and warns against over-interpreting PTR data.
  • IETF IP Geolocation Workshop materials: workshop context for treating network-provided location signals as scoped, fallible hints rather than ground truth.
  • CAIDA Hoiho / ITDK: learns operator-specific regular expressions from router hostnames for infrastructure/geolocation inference; this project borrows its explainable-regex philosophy.
  • DRoP, DNS-based Router Positioning: prior work on extracting and validating router location hints from hostnames.
  • AWS EC2 public hostname documentation: documents the ec2-A-B-C-D.<region>.compute.amazonaws.com form used by the provider-specific rules.

The classifier emits labels in MISP machine-tag form (ptrclassify:predicate="value"). A MISP taxonomy definition suitable for validation or import is provided in misp-taxonomy/machinetag.json and officially published as a MISP taxonomy.

The built-in taxonomy is not claimed to be an Internet standard. It is designed as a practical, extensible CTI/network-enrichment taxonomy with orthogonal namespaces instead of a single mutually-exclusive class.

Publishing

Releases are published by the Publish to PyPI GitHub Actions workflow. Before the first release, configure a PyPI trusted publisher for this repository with workflow name publish.yml and environment name pypi; no long-lived PyPI API token is required.

To publish a new version, update project.version in pyproject.toml, merge the change, and publish a GitHub release. The workflow builds both the source and wheel distributions, validates them, and publishes them to the ptrclassify PyPI project. It can also be started manually from the Actions tab when a release job needs to be retried.