NTC Atlas

An interactive, evidence-backed atlas of network traffic classification research.

Live site: https://ntc-atlas.pages.dev

NTC Atlas follows the evolution of network traffic classification from port rules and flow statistics to deep learning, pretrained models and traffic foundation models. Encrypted traffic is a central theme. Every analytical statement links to its source, and paper claims, later criticism and curator interpretation stay separate. The atlas is curated and maintained by Mohamed Hashim Changrampadi.

Papers	Authors	Datasets	Relationships	Evidence records
93	318	19	107	641

Who it is for

• PhD students entering traffic classification and looking for a reading path.
• Researchers checking which results are comparable, which datasets are overused and where leakage may exist.
• Reviewers and practitioners asking what is reported about throughput, latency, memory and deployability.

Features

• Timeline of methodological generations with collision-aware labels.
• Connections graph of documented extensions, critiques and baseline comparisons. Each relationship type is drawn differently, and a baseline comparison is never shown as a successor.
• Techniques view of how inputs and model families changed.
• Compare up to five papers side by side. Each value keeps its unit and scope, and missing values carry a qualified NA.
• Datasets with label provenance, assignment units and version caveats.
• Authors and institutions built from verified publication bylines, with confirmed Google Scholar profiles where known.
• Research-gap matrix across 16 dimensions, each cell linked to evidence.
• Research briefings (coming soon): curator analysis of where the field is heading, published as a newsletter.
• Field history and a full methodology page.
• Search, filters, sorting and pagination. Every view has a shareable URL, e.g. ?view=authors&person=luming-yang.
• Light and dark themes.
• Exports of any selection as CSV, JSON or BibTeX, plus the raw JSON corpus.

Data and evidence methodology

The corpus lives in versioned JSON files in public/data/:

File	Content
papers.json	Bibliography, methods, evaluation fields, gap states and review depth
authors.json	Names, aliases, dated affiliations, papers and profile links
datasets.json	Provenance, traffic units, labels, caveats and recorded uses
relationships.json	Directed paper-to-paper relationships with supporting evidence IDs
evidence.json	Source URL and type, paraphrase, locator, verification status and access date
label-provenance.json	Label evidence, assignment unit and semantic object per dataset
history.json	Evidence-linked narrative periods
meta.json	Scope, counts, author-inclusion rule and evidence boundaries

Curation rules:

• Primary sources first. Titles, author order, year and venue are checked against the original paper, DOI record, publisher, proceedings or authoritative repository. Preprints and peer-reviewed versions are tracked separately.
• Authors enter only through verified bylines. Seed lists and Google Scholar are used for discovery and identity. A publication byline and a dated affiliation are verified separately. Committee membership is never evidence of a relevant publication.
• Relationships need evidence. An extension, critique or comparison is recorded only when a source states it. Chronology alone never creates an edge.
• Qualified NA. Comparison fields say NA: not reported in reviewed sections, NA: not applicable or NA: not yet extracted.
• Scoped results. Accuracy and F1 appear with their dataset, split, task and metric. Offline samples per second are never presented as live network throughput.
• No reproduction claims. Linked artefacts have not been rerun.

The Methodology view on the site explains these rules in full.

Data limitations

This is a curated, evolving corpus, not a systematic review. Technical extraction depth varies by paper, and each record states its review depth. A gap cell marked "Not addressed/not reported" means the reviewed evidence did not establish coverage. Absence from this corpus is not evidence of absence from the field. Open extraction work is listed in docs/RESEARCH-BACKLOG.md.

Run locally

Requires Node.js 22 (see .node-version) and Corepack, which ships with Node.

git clone https://github.com/cmhashim/ntc-atlas.git
cd ntc-atlas
corepack enable
pnpm install --frozen-lockfile
pnpm dev            # http://localhost:5173

The project pins pnpm through the packageManager field in package.json.

Validate and build

pnpm validate       # corpus integrity checks (scripts/validate-data.mjs)
pnpm typecheck      # TypeScript
pnpm lint           # ESLint
pnpm build          # static export into out/
pnpm check          # all four, as CI runs them
pnpm preview        # serve out/ locally through Wrangler (http://localhost:8788)

The build is a fully static site. All views run in the browser over the bundled JSON, so hosting needs no server, database or API key. To host under a sub-path, e.g. GitHub Pages, build with NEXT_PUBLIC_BASE_PATH=/ntc-atlas pnpm build.

Deployment

.github/workflows/deploy.yml runs the four checks on every push and pull request.

• Pushes to main deploy to Cloudflare Pages at https://ntc-atlas.pages.dev.
• Pull requests from this repository get a preview URL in the workflow summary.
• A manual deploy is available under Actions → Check and deploy → Run workflow.

The workflow needs two repository secrets: CLOUDFLARE_API_TOKEN (scoped to Account → Cloudflare Pages → Edit) and CLOUDFLARE_ACCOUNT_ID.

Updating the corpus

1. Verify the primary source before adding a paper to scripts/curate.py. Put unresolved candidates in the pending queue.
2. Resolve duplicates, versions and author aliases. Keep distinct follow-up models separate.
3. Record analytical evidence with URL, source type, paraphrase, locator, verification status and access date.
4. Add only relationships supported by their own evidence.
5. Scope every gap judgement to reviewed evidence and explain it in the cell note.
6. Regenerate, then run the checks:

python3 scripts/curate.py
pnpm counts         # refresh the counts table above
pnpm check

Keep history.json curated by hand. Change the research date in meta.json only after a renewed source review.

Citing

If you use the atlas or its data, please cite it. GitHub's Cite this repository button (from CITATION.cff) gives APA and BibTeX formats.

Changrampadi, M. H. NTC Atlas: an interactive, evidence-backed atlas of network traffic classification research. https://github.com/cmhashim/ntc-atlas

Cite the original papers for any result you take from the atlas.

Contributing

Suggestions for papers, metadata corrections and author identity fixes are welcome. Please use the issue forms and read CONTRIBUTING.md first. Every change needs a primary source.

Maintainers: see docs/MAINTENANCE.md for the weekly discovery review and recurring upkeep.

Licence

• Code: MIT, see LICENSE.
• Curated data in public/data/: CC BY 4.0, see LICENSE-DATA.
• Vendored third-party files keep their own licences (vendor/).
• The cited papers, datasets and web pages belong to their authors and publishers. The atlas stores citations and short paraphrases only, not PDFs or traffic captures.

Planned work is in docs/ROADMAP.md. Earlier changes are in CHANGELOG.md. Browser QA notes from earlier releases are in docs/QA-HISTORY.md.