PGPO: Privacy-Preserving Recommendations via Prototype-Guided Obfuscation
August 25, 2026
PGPO protects LLM-enhanced recommenders from inversion attacks by using GRPO to learn topology-aware obfuscation for high-influence prototype words. The method shifts semantic embeddings to protect privacy while preserving the relative item relations needed for utility.
HOW THIS AFFECTS YOU
●
builderYou can protect user data in recommendation systems without breaking the accuracy of your model.
●
policyThis provides a technical mechanism for implementing privacy-preserving semantic embeddings.