ZJU's PGPO obfuscates embeddings to defend LLM recommender inversion attacks

zju · hf · 2026-11-19

Zhejiang University proposes PGPO (Prototype-Guided Progressive Obfuscation) to defend LLM-enhanced recommender systems against embedding inversion attacks. It identifies high-influence prototype words, uses GRPO-guided LLM variant exploration and anchor-oriented graph expansion, with rewards balancing embedding dissimilarity, structural consistency, and semantic constraints — perturbing high-influence semantic embeddings while preserving relative item relations.

Original post →

More from Safety

Safety channel →