Activation Steering Meets Preference Optimization: Defense Against Jailbreaks in Vision Language Models

Documenti analoghi

• Steering Away from Harm: An Adaptive Approach to Defending Vision Language Model Against Jailbreaks
  di: Wang, Han, et al.
  Pubblicazione: (2024)
• Adversarial Defense in Vision-Language Models: An Overview
  di: Fu, Xiaowei, et al.
  Pubblicazione: (2026)
• Defense-to-Attack: Bypassing Weak Defenses Enables Stronger Jailbreaks in Vision-Language Models
  di: Zhao, Yunhan, et al.
  Pubblicazione: (2025)
• TAIJI: Textual Anchoring for Immunizing Jailbreak Images in Vision Language Models
  di: Yin, Xiangyu, et al.
  Pubblicazione: (2025)
• White-box Multimodal Jailbreaks Against Large Vision-Language Models
  di: Wang, Ruofan, et al.
  Pubblicazione: (2024)