LaCache: A New Semantic Caching Scheme to Thwart Cache-Collision Attacks in LLM Serving

A new research paper on arXiv (2608.01718) introduces LaCache, a semantic caching scheme designed to defend against cache-collision attacks in large language model (LLM) serving. Semantic caching reuses responses to semantically similar requests via embeddings, offering faster responses and reduced costs, but existing schemes are vulnerable to cache-collision attacks where an adversary pollutes the cache with crafted queries, corrupting responses to subsequent legitimate requests. LaCache addresses this by checking not only the cache hit of the query but also the cache hit of its first k speculatively decoded tokens. This design provides formally guaranteed resilience against such attacks. The paper is authored by researchers and was announced as a new submission on arXiv.