Catching the never-seen: novelty for zero-days

“Zero-day” is an overused word. Made concrete, it means traffic unlike anything a given deployment normally emits. A supervised classifier, however good, only recognises the families it was trained on. To catch the rest, you need a judge that models normal rather than known-bad.

Modeling normal, per tenant

Ward³'s novelty judge learns the benign manifold of each tenant and scores how far a flow's morphology sits from it. The score is stratified per tenant — one tenant's normal is not another's — and, like every judge, it passes through conformal calibration before it means anything.

Masked reconstruction, on purpose

The mechanism is deliberately simple and inspectable, not a black box:

• It reconstructs the runtime feature vector from a low-dimensional model of benign traffic.
• The error is masked over the features actually observed, so missing signals never fake a high score.
• A minimum number of observed features is required, or the judge abstains rather than guess.
• The raw error never leaves bare — it becomes a conformal p-value first.

Observation first, corroboration always

In its first form the novelty judge only observes: it enriches events and feeds the mediator, but never blocks on its own. A novelty detector is a detector of the unusual, not a detector of the malicious — so it corroborates, it does not decide alone.

Measured on held-out attack families it was never trained on, it recovers detections the supervised judges miss. That is the whole point: coverage the trained models cannot have, added without loosening the false-positive bound.