Notes on adversarial-grade detection
Engineering write-ups, research notes and talks on building an NDR that assumes attackers target the machine-learning model too.
A companion to the Ward³ whitepaper: what has to be true for a product to honestly call itself an Adversarial Neural Mediator — and why three of the five criteria are the ones vendors quietly skip.
Most NDR products bet that one well-trained model can catch advanced attacks. Open-source evasion tooling quietly turned that bet into a liability.
“xx% fewer false positives” is an average measured on someone's benchmark. A conformal bound is a property that holds on your traffic.
How the mediator turns three opinions into one verdict — and why the moments they disagree are the ones worth watching.
Supervised judges catch the attack families they were shown. A novelty judge catches the flows that resemble nothing the deployment normally produces.
A conference talk on treating the ML model as an attack surface, and what a network detection stack looks like once it does.