Concept · 1 episode(s)

Adversarial Fine-Tuning

← all concepts

Definition

Adversarial Fine-Tuning is the practice of retraining a model on examples of attacks that previously succeeded against it, in order to patch those specific vulnerabilities. It is a common defensive countermeasure in security-style ML competitions, but it can backfire by overfitting the defender to last round's tricks while leaving it blind to the next novel attack — or even degrading its performance on ordinary, non-adversarial inputs.

Episodes covering this