Ethics, licences and data handling

DriftWatch studies how malware detectors decay over time, and what label delay, a small analyst budget and drift monitoring do to that decay. It does not study how to defeat detectors.

What this project never does

  • Never downloads, runs, opens, decompiles, unpacks or redistributes malware samples or any binary. The project uses pre-extracted feature files only: LAMDA’s Drebin-style vectors, BODMAS’s EMBER-style vectors and, optionally, EMBER2024 vectors.

  • Never builds evasion tooling and never generates adversarial examples. It publishes no detector tuned for evasion and no guidance on bypassing detection.

  • Never commits dataset content. data/ and results/ are git-ignored. The only data in the repository is a synthetic fixture under 1 MB with the same schema (Phase 2), created for tests and CI.

Licences

Item

Licence

Use here

DriftWatch code

MIT (ADR-0002)

This repository

LAMDA

MIT [verified in build guide §1]

Feature files; cite the dataset paper; not redistributed

BODMAS

The repository code is BSD-2-Clause; no licence is stated for the dataset files (data card, checked 20 Oct 2026; the build guide’s BSD-2-Clause for the dataset is [verify])

Feature files; cite; not redistributed (downloaded into the git-ignored data/ only)

EMBER2024 (optional, Tier 3)

Apache-2.0 [verified in build guide §1]

Win32 weekly subset only; not redistributed

AndroZoo

Academic key; redistribution forbidden

Not used (would need a faculty-sponsored key)

SOREL-20M

Own terms, not verified

Excluded (build guide §10)

Every dataset version, source URL and file hash goes into data/manifest.json (Phase 2) and its data card under docs/data_cards/. The full licence audit is docs/licence_audit.md (22 Oct 2026; regenerate the dependency table with uv run python scripts/list_licences.py).

Data handling

  • Raw files and caches live under data/ on the researcher’s machine; results live under results/. Both are git-ignored, and the guard hook blocks git add of either (except data/manifest.json).

  • Claude Code permissions deny editing data/** and reading secrets (.env, secrets/, keys).

  • Labels come from the datasets’ own labelling (antivirus-based). LAMDA’s card warns of antivirus-centric labelling bias; data cards and limitations will say so (plan Appendix E, threats 20–21).

  • No personal data is collected. The feature files describe software, not people.

Honest reporting

  • Every number of consequence goes into docs/CLAIMS.md, with its file and commit, before it is written anywhere else.

  • Hypotheses and analysis rules are pre-registered (docs/PREREG.md, locked by git tags) before results exist. Negative and flat results are reported as they are.

  • No contribution is claimed as “first”; the text states what was measured.

AI assistance

Claude Code (an AI coding agent) implements, tests and drafts under the researcher’s direction. The researcher decides scope, approves every pre-registration, ADR and gate, verifies each statistic and rewrites all report and paper text in their own words.

Item

Status

VISTAS AI-assistance disclosure policy

unknown, to verify (researcher’s answer, 10 Oct 2026)

Disclosure plan until verified

Disclose AI assistance in the report and the paper (plan default)

Other VISTAS rules (deadline, template, page rules, demo)

unknown, to verify; tracked in docs/PHASES.md