LCGuard: Safe KV Cache Sharing in Multi-Agent LLM Systems
A novel framework named LCGuard (Latent Communication Guard) has been developed to promote secure KV-based latent communication in multi-agent systems utilizing large language models (LLMs). Although many current multi-agent systems rely on natural language for communication, recent studies indicate that using transformer key-value (KV) caches for latent communication can enhance efficiency and retain more relevant task information. Nonetheless, these KV caches can also contain contextual inputs, intermediate reasoning states, and agent-specific details, leading to a potential risk of sensitive information leaking between agents without direct textual indication. LCGuard treats the shared KV caches as latent working memory and learns to transform representations before these cache artifacts are exchanged. This framework aims to reduce privacy concerns and was published on arXiv under ID 2605.22786.
Key facts
- LCGuard is a framework for safe KV-based latent communication in multi-agent LLM systems.
- It addresses privacy risks from KV caches encoding sensitive content.
- KV caches improve efficiency and preserve task-relevant information.
- LCGuard learns representation-level transformations before cache sharing.
- The paper is available on arXiv under ID 2605.22786.
- Multi-agent systems increasingly rely on intermediate communication.
- Latent communication through KV caches is an emerging approach.
- LCGuard formalizes representation-level transformations.
Entities
Institutions
- arXiv