SelfDefend: LLMs Can Defend Themselves against Jailbreaking in...
summary
SelfDefend implements a shadow LLM that runs in parallel to a target model to detect harmful intent, using two detection templates (direct and intent prompts) and checkpoint-based access control; the team distills GPT‑4 red‑team data and fine‑tunes Llama‑2 (Aurora) as a low‑cost defense, deploys against GPT‑3.5/4, Claude, Llama‑2 and Mistral, outperforms seven prior defenses, matches GPT‑4 SelfDefend, shows >80% normal queries with no added latency and a max 0.4s extra delay, and resists adaptive jailbreaks.
AI-written summary · about
chapters (8)
- 0:00SelfDefend framework
- 2:51Motivations behind SelfDefend
- 3:57SelfDefend pipeline and prompts
- 6:26Evaluation on GPT models
- 7:50Fine-tuning open-source models
- 8:31Performance comparison
- 10:44Robustness against adaptive jailbreaks
- 11:50Conclusion and future work
related talks
01h 04mRobust Generalization in the Era of LLMs: Jailbreaking Attacks and Defenses
37mSecure by Design: Strategies for LLM Adoption in Cloud-Native Environm...
59mRed Teaming Large Language Models
28mJailbreaking and Securing LLM Apps: Lessons from an online wargame experiment
46mGuarding the LLM Galaxy: Security, Privacy, and Guardrails in the...
26mOne guardrail is not enough: Defense-in-depth against LLM jailbreaks
09mLLM hacking is underrated
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges