An MLOps project for detecting network attacks with XGBoost and simulating continual retraining on UNSW-NB15 traffic. Zeek supplies live connection logs, DVC versions datasets and exported models, and MLflow records experiments and model versions.
- Offline UNSW-NB15 preprocessing and chronological stream batches.
- Live Zeek capture, ingestion, and mapping into the training feature schema.
- Reproducible training and validation with Recall, F1, and PR-AUC gates.
- GitHub Actions unit tests and a training pipeline that registers passing
candidates in Staging with the
challengeralias. - Manual Production promotion, model export, and inference verification.
Scheduled retraining, drift detection, and ground-truth labeling remain planned extensions. Automatically assigned labels on live logs require human verification before those records can support model evaluation.
Use Python 3.11 and run commands from the repository root:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements-test.txt
python -m pytest -q --junitxml=reports/pytest.xmlUnit tests use small fixtures and mocks; they run without the full datasets, Docker, Zeek, or a running MLflow service. See testing details.
For training and registry operations, install requirements-automation.txt instead.
requirements.txt contains the full pinned environment for exploration and
notebooks; the smaller profiles use its versions as constraints.
flowchart LR
A[Versioned dataset] --> B[Validate data and split]
B --> C[Train XGBoost candidate]
C --> D[Evaluate Recall, F1, PR-AUC]
D --> E{All metrics >= 0.95?}
E -->|Yes| F[MLflow Staging / challenger]
E -->|No| G[Fail pipeline and retain current model]
F --> H[Manual Production promotion]
The automatic gate reserves 20% of flow groups from dataset.csv for validation,
using seed 42. Duplicate flows stay together in one partition. Batch 7 remains
reserved for final testing and is excluded from the automatic gate.
Pushes and pull requests targeting main run tests on GitHub-hosted runners.
Successful pushes to main, or manual runs on main, continue on a self-hosted
runner with access to the DVC remote and persistent MLflow storage.
| Path | Purpose |
|---|---|
src/production/ |
Capture, ingestion, preprocessing, training, evaluation, and registry CLIs. |
src/simulation/ |
Offline cleaning and historical stream simulation. |
src/config.py, src/dataset_io.py |
Shared configuration and dataset storage helpers. |
tests/ |
Pytest suite and detailed testing notes. |
notebook/ |
Exploratory analysis and preprocessing experiments. |
data/ |
Raw and processed datasets; large files are managed by DVC. |
models/ |
Exported Production model pointer and metadata. |
metrics/ |
Training metrics tracked by Git. |
zeek_logs/example/ |
Sample capture logs for ingestion examples. |
docs/ |
Operations guide. |
.github/workflows/ |
Automated testing and training workflow. |
Training settings live in params.yaml, pipeline stages in dvc.yaml, and local S3 storage in docker-compose.yml. Local datasets, captures, reports, credentials, and MLflow storage are excluded from Git.
- Operations guide: setup, DVC, simulation, live capture, configuration, Actions runner setup, and model registry commands.
- Testing guide: test coverage and local execution.
The recorded Production model is adaptive-nids-xgb v1 (champion), trained on
dataset v1.1.0. Its exported identity and lineage are recorded in
models/production_model.yaml; see the
registry reference for usage
and historical evaluation results.
Author: Alexander Angelo (@xx-bot). Course: MLOps, Semester 5. Supervisor: Rizal Setya Perdana, S.Kom., M.Kom., Ph.D. Licensed under the MIT License.