Skip to content
Tracks @XLab
Sign in

Defer to resample, mechanically

Defer to resample, mechanically

innocent step — min of 3 samplesattacked step — min of 3
defer →audit (top 1% of innocent minima) →more suspicious →

Win — step audited

70%

Lose — attack executed, unaudited

30%

Continue — attack fizzles

0.0%

Innocent step deferred to T

0.8%

Schematic distributions, not the paper's data. The outcomes are win/continue/lose for one attacked step. Raise k with c at 100%: innocent minima drift left (fewer wasted audits and deferrals) while the attacked step's minimum stays suspicious — the win:lose odds improve. Now lower c: attack steps mostly fizzle into "continue" (selection for benign actions), and every incomplete attack is one more chance to catch the model on a later step.

Embed this demo

Use the chrome-less embed view inside an iframe:

<iframe src="/demos/defer-to-resample/embed" width="100%" height="360" style="border:0"></iframe>