Definition
Plain language
Retraining a detector on the exact tricks that fooled it, so it learns to catch them next time.
As stated in the literature
Fine-tuning a classifier on adversarial examples that previously evaded it, hardening it against a known attack distribution; effective only against attacks that leave a consistent, learnable signature.
Why it matters: It lets a detector close known loopholes instead of being fooled by the same trick over and over, though only for attacks that repeat a recognizable pattern.
For example, after spam messages slip past a filter by misspelling words, the team feeds those same tricky messages back in so the filter learns to flag them.
Heard on the show
“That's adversarial fine-tuning.”Episode 220 — Write Like It's 1923: The One-Prompt Trick That Beats AI Detectors