LCGuard Framework Addresses Privacy Risks in Multi-Agent AI Systems Using Latent Communication
Researchers have introduced LCGuard (Latent Communication Guard), a framework designed to secure key-value (KV) cache sharing in multi-agent large language model (LLM) systems, according to a paper published on arxiv.org.
According to the research, while multi-agent LLM systems increasingly use latent communication through transformer KV caches for efficiency, these caches “encode contextual inputs, intermediate reasoning states, and agent-specific information, creating an opaque channel through which sensitive content may propagate across agents without explicit textual disclosure.”
LCGuard addresses this by treating shared KV caches as latent working memory and learning representation-level transformations before cache artifacts are transmitted between agents. The framework defines unsafe cache artifacts as those from which “an adversarial decoder can recover agent-specific sensitive inputs,” according to the paper.
The system uses adversarial training where “the adversary learns to reconstruct sensitive inputs, while LCGuard learns transformations that preserve task-relevant semantics and reduce reconstructable information,” the researchers state.
According to arxiv.org, empirical evaluations across multiple model families and multi-agent benchmarks show that “LCGuard consistently reduces reconstruction-based leakage and attack success rates while maintaining competitive task performance compared to standard KV-sharing baselines.”
The research appears alongside other work exploring latent communication in multi-agent systems, including LACO for collaborative driving, which also addresses latency and information challenges in agent coordination.