Talk — Adversarial Neural Mediation
A conference talk on treating the ML model as an attack surface, and what a network detection stack looks like once it does.
Most detection talks assume the model is a neutral observer. This one starts from the opposite premise: the model is a target, and the interesting engineering begins once you accept that.
What the talk covers
Roughly twenty-five minutes, in five moves:
- The thesis behind Adversarial Neural Mediation (ANM), and why the threat model shifted.
- Why architectural orthogonality beats ensembling under gradient attack.
- The mediator, and divergence used as a security signal rather than a footnote.
- Conformal calibration as a way to bound false positives, not just tune them.
- The tamper-evident, post-quantum audit ledger that makes every verdict replayable.
The through-line
The talk closes on the five criteria a product has to meet to honestly call itself ANM, using Ward³ as the reference implementation — and on the one question it keeps coming back to: not whether an attacker will test your ML, but what catches the attack when one model gets it wrong.
Who it's for
SOC leads and detection engineers deciding how much to trust a single model, and ML-security researchers interested in defence rather than only attack. Slides and recording are available on request.
Ward³ is the reference ANM implementation. If you defend a regulated environment, we can look at where it fits.
Request access