AI Grimoire
06 — siblings
evasion/
poisoning/privacy/stealing/agents/

Adversarial ML / evasion

Evasion

Gradient-based search for an input that produces a chosen output. The continuous case is well understood; the discrete case — searching over tokens rather than pixels — is where the current attacks on language models live.

2 entries.

Entries

06.01.1
Greedy Coordinate Gradientstandard
Discrete optimisation of an adversarial suffix against a refusal-trained model.
O(k·B·d)
06.01.2
Projected Gradient Descent Attackstandard
Iterative L_p-bounded perturbation maximising task loss.
O(k·d)