Similarity Is Not Validity: Defending LLM Semantic Caches Against Poisoning
This work proposes a novel defense that recovers necessary information needed to distinguish valid from invalid cache hits from the raw text of the cache key, and proves that Deletion Gain stays positive when a deletion leaves text close enough to the rewrite, and searches for such deletions with a sliding window.
Zi-Han Zhang, Shuang-Jie Yao, Ze-Sen Liu et al.
· 0 citations