What is LCGuard and how does it work?

LCGuard is a security framework that protects multi-agent LLM systems by transforming shared KV cache representations to block sensitive content transmission without degrading task efficiency.

Why is KV cache sharing a privacy risk?

Shared KV caches encode context and reasoning states in opaque vectors, creating covert channels that silently leak proprietary or private data between agents.

How effective is LCGuard and where is it heading?

Experiments show LCGuard cuts reconstruction attacks while maintaining performance, enabling secure AI agent deployment in sensitive sectors like finance.

LCGuard：多智能體系統中基於潛在通信的安全KV快取共享框架

本文針對基於大語言模型的多智能體系統中，利用Transformer鍵值（KV）快取進行潛在通信所引發的隱私洩露風險，提出了LCGuard框架。現有研究指出，儘管KV快取共享能提升效率並保留豐富資訊，但其編碼了上下文輸入與推論狀態，成為敏感資訊傳播的隱蔽通道。LCGuard將共享KV快取視為潛在的工作記憶，透過在學習表示層施加變換來阻斷敏感內容的傳輸。該方法形式化了基於重構的表示層敏感資訊洩露問題——若對抗解碼器能從快取中恢復敏感輸入，則判定為不安全。透過對抗訓練，LCGuard學習保留任務語義並最小化可重構資訊，同時對抗者學習重構敏感輸入。在多個模型家族與多智能體基準上的實驗表明，LCGuard在維持具競爭力的任務效能的同時，顯著降低了基於重構的洩露與攻擊成功率，為多智能體系統的安全協作提供了新範式。