Skip to content

PGPO: prototype-guided progressive obfuscation for privacy-preserving LLM-enhanced recommendation

PGPO: Prototype-Guided Progressive Obfuscation for Privacy-Preserving LLM-Enhanced Recommendation

arXiv: 其他 AI 安全论文· arXiv:2611.23236· Shengxiang Lin, Jiajie Su, Pengyang Zhou, Xiang Chen, Xiaolin Zheng, Feng Tian, Chaochao Chen·Indexed by this site · 13 hr ago· Originally publishedDefenses· Its place in the scan record →
AI summary

The paper proposes PGPO, which uses prototype-guided progressive obfuscation to prevent item semantic embeddings in LLM-enhanced recommendation from leaking sensitive text attributes through inversion attacks. The method learns topology-aware obfuscation anchors from high-impact prototype words and propagates them over a static semantic graph, optimizing with GRPO under rewards for individual difference, structural consistency and semantic constraints. Experiments on two benchmarks show strong inversion defense, with recommendation performance close to the upper bound.

Read originalarxiv.org