Definition
Plain language
A defense that runs an AI's math at higher numerical precision, which erases the rounding quirks a hardware backdoor relies on.
As stated in the literature
An inference technique that stabilizes floating-point rounding by computing in higher precision; collapses the FloatDoor hardware-fingerprint channel from near-100% to under 1% with negligible accuracy loss, at a runtime and memory cost.
Why it matters: It neutralizes a platform-triggered backdoor almost entirely with little loss in accuracy, offering a practical defense at some extra runtime cost.
For example, running the model's math more carefully erases the tiny rounding quirks that a hardware backdoor was secretly listening for.
Heard on the show
“The main one is running inference in higher numerical precision — they call it LAYERCAST.”Episode 158 — How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave