NonTextual Target Attack (NTA) optimizes for maximized unsafety probability rather than fixed response patterns, bypassing the constraints of traditional gradient-based attacks. It decomposes the non-textual objective into two differentiable unconstrained losses for efficient optimization.
HOW THIS AFFECTS YOU
●
researcherThis demonstrates a more efficient gradient-based adversarial method that expands the attack search space.
●
policyYou must account for attacks that do not rely on predictable text patterns to evaluate model safety.