LaCache: Robust semantic caching for LLM services

The authors propose LaCache, a redesign of semantic caching for LLM services, to defend against cache collision attacks. Its core idea is that although an attacker can fully control the injected query, they have limited control over the model-generated response; therefore LaCache moves the integrity check from the query to the response, additionally checking whether the first k speculative decoding tokens of the query hit the cache. The authors' analysis states that the method can detect, with high probability, cache collision attacks that produce semantically different responses. Evaluations on multiple LLMs and benchmarks show that the method reduces the attack hit rate to near zero while retaining more than 90% of benign cache utility.