CachePrune: Privacy-Aware KV Cache Sharing Prevents Side-Channel Attacks While Preserving 85%+ Reuse

arXiv·medium signal

Unrestricted cross-user KV cache sharing in LLM serving systems creates side-channel vulnerabilities where adversaries can infer user inputs by probing for cache reuse. CachePrune introduces fine-grained, privacy-aware sharing that prevents leakage without disabling sharing entirely. Existing defenses are all-or-nothing; CachePrune preserves substantial reuse potential while closing the side-channel — critical for multi-tenant LLM serving deployments.