Research question
Can randomised smoothing ideas defend against jailbreak suffixes?
Key findings
- Perturbing input copies and aggregating responses reduces GCG success below 1% - Robust to adaptive attacks in the paper's threat model
SmoothLLM
Applied randomised-smoothing intuition to LLM inputs, a lightweight defence against character-level adversarial suffixes.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.