ADR 0004: scikit-learn MLP and autoencoder instead of PyTorch

• Status: accepted
• Date: 2026-09-26

Context

The spec lists PyTorch for the autoencoder and adversarial training. The models here are tiny (thousands of parameters on 15-22 features); the development machine and CI are CPU-only; PyTorch adds hundreds of MB to every CI install and lags new Python releases.

Decision

Use sklearn.neural_network.MLPClassifier for the supervised member and MLPRegressor trained X -> X on benign flows as the autoencoder. White-box gradients for the MLP are computed analytically from coefs_ / intercepts_ (MLPDetector.grad_z), and a unit test checks them against central finite differences to 1e-4.

Consequences

• Fast CI, deterministic seeds, trivial serialisation with joblib.
• No GPU training, which is fine at this scale: the full CIC-IDS2017 study runs on a laptop CPU.
• If larger models are ever needed, only grad_z and fit_z must be re-implemented behind the same interface.