Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog and the project uses Semantic Versioning.

[Unreleased]

Repository renamed to rakshit-737/feint-adversarial-ids: repo, docs (https://rakshit-737.github.io/feint-adversarial-ids/), badge, citation and package URLs updated; the container image is now ghcr.io/rakshit-737/feint-adversarial-ids. Older entries below keep the pre-rename names (rakshit-737/feint, ghcr.io/rakshit-737/feint).

Repository renamed to rakshit-737/feint-adversarial-ids. All links, badges, CITATION.cff, mkdocs, pyproject URLs and the paper now point to https://github.com/rakshit-737/feint-adversarial-ids and https://rakshit-737.github.io/feint-adversarial-ids/; the container image is now published as ghcr.io/rakshit-737/feint-adversarial-ids. Older entries below keep the old names as history.

Moustafa & Slay 2016 DT figures (85.56 % / FAR 15.78 %) now marked unverified in README, paper and reproduction output: no legitimate open copy or open secondary source restating them was found.

Datasets: CSE-CIC-IDS2018 (three days used; all ten SHA-256 pinned from CI downloads whose size and S3 multipart ETag were verified, download_cicids2018.py --print-pins), corrected CIC-IDS2017 (Engelen et al. 2021 / Liu et al. 2022), CICIoT2023 ipfixprobe subsample.
feint xdata cross-dataset study (corrected CIC-IDS2017 <-> CSE-CIC-IDS2018, 5 seeds).
feint ablate: paired schema ablations over 5 seeds (kNN sanitiser in robust-feature space vs all features; adversarial training on schema-constrained vs textbook examples).
feint repro: Moustafa & Slay 2016 and Sharafaldin et al. 2018 reproductions, the latter also on the corrected CIC-IDS2017 and with a regularised-QDA variant labelled as a deviation.
A2PM reproduction (Vitorino et al. 2022), scripts/repro_a2pm.py / bench study repro-a2pm on Python 3.11: the paper's RF and Keras MLP, 50-iteration targeted and untargeted attacks with scores after every iteration. a2pm's per-value loops are replaced by vectorised pattern subclasses that every run checks against the library (transform level and end to end). A run with the library's own transforms is kept as a cross-check.
Dashboard screenshot in the README hero and on the docs home page.
feint seeds --adv --poison: adversarial training and poisoning with 5-seed CIs.
bench.yml workflow for every heavy study; paper.yml builds the preprint draft in paper/.
Docs: Evaluation and Reproduce pages, dashboard over all datasets; CITATION.cff, templates.

Headline numbers are now 5-seed means with CIs (they moved by up to 0.07 from the seed-0 table).
Hardened tree models are attacked through both the undefended and the hardened MLP.
Corrected claims: budget semantics (constrained and textbook attacks are not budget-matched), defender-favourable schema assumptions, UNSW derived-feature agreement, Sharafaldin setup.
Docstrings and full type hints on every public function and method of the feint package (109 of 109); ruff now enforces pydocstyle (Google convention) and flake8-annotations on feint/.
bench.yml study steps fail on a failing pipeline (pipefail).

Download helper: single stream, per-destination lock, size and hash checks; a corrupted local CSE-CIC-IDS2018 file produced by two concurrent downloaders was deleted and never used.
API input validation, data root default, single-threaded RF scoring, release notes extraction, sdist fixtures.

feint seeds: multi-seed robustness study with 95 % Student-t CIs (and Wilson intervals); 3-seed results for CIC-IDS2017 and UNSW-NB15 in results/*/seeds.{json,md}.
feint steal: model-stealing study (label-only queries -> transfer attack); UNSW-NB15 results.
MkDocs Material docs site on GitHub Pages with a static results dashboard (/demo/).
Dockerfile (slim, non-root), docker-compose for the scoring API, release workflow publishing ghcr.io/rakshit-737/feint and wheel/sdist on v* tags.

Real data: CIC-IDS2017 (MachineLearningCVE, 2,830,743 flows) and UNSW-NB15 (official 175,341 / 82,332 partition) loaders, with SHA-256-verified, resumable download scripts.
Declarative domain-constraint schemas (feint/schema.py) for both datasets: monotone, range-bounded, integer, fixed and derived features plus cross-feature relations.
Detector ensemble: XGBoost + MLP (supervised) OR a benign-trained autoencoder (unsupervised); logistic regression, random forest and IsolationForest baselines.
Attacks: white-box PGD, score-based black-box random search, and an adaptive attack (transfer from a differentiable surrogate + black-box refinement) for trees and the ensemble. Unconstrained reference variants.
Hardening: generic adversarial training (MLP and XGBoost), robust-feature models, an adversarial-input detector.
Explanations: exact TreeSHAP (XGBoost native) and constraint-valid sparse counterfactuals.
Poisoning: backdoor study with an attacker-controllable trigger and a robust-feature kNN training-set sanitiser.
Temporal drift study on CIC-IDS2017 (train Monday-Wednesday, test Thursday-Friday).
Markdown + PNG reports; FastAPI scoring/explanation service (feint serve).
Real benchmark results in results/cicids2017/ and results/unsw_nb15/.
LICENSE (MIT), CONTRIBUTING, CHANGELOG, ADRs in docs/adr/.

Robustness curves are monotone by construction (an attacker with budget eps may use any smaller budget's successful flow).
Byte bounds follow payload semantics (CIC) and header-inclusive semantics (UNSW).
CI runs ruff, the test suite, and a quick end-to-end smoke study.

The synthetic-only results/report.* from 0.1.0 (superseded by real-data results).

MVP on synthetic flows: MLP detector with analytic gradients, constrained PGD, adversarial training, IsolationForest backstop, base-rate precision, CLI, CI.