Reproduce “Contextualized Privacy Defense for LLM Agents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yule Wen, Yanzhe Zhang, Jianxun Lian, Xiaoyuan Yi, Xing Xie, Diyi Yang

Researchers from Tsinghua, Georgia Tech, Microsoft, and Stanford introduce Contextualized Defense Instructing (CDI), an experience-driven reinforcement learning framework designed to enhance Large Language Model (LLM) agents' contextual privacy awareness. Optimized CDI achieves 94.2% privacy preservation and 80.6% helpfulness on unseen scenarios, demonstrating superior robustness against adaptive adversarial attacks while balancing utility.

Key insights

  • Integrating a dedicated instructor model to provide proactive, context-aware guidance after tool-call results is more effective for privacy defense than static prompting or reactive guarding.
  • Learning from adversarial failure cases through a reinforcement learning framework enables defenses to develop more fundamental and adaptive privacy reasoning, leading to superior robustness and generalization against strategic attacks.
  • The optimized CDI approach significantly improves the privacy-helpfulness trade-off and can empower even weaker agent backbone models to achieve high privacy performance by offloading complex privacy reasoning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.