Using Large Language Models to protect information search in Multi-Domain Operations

We consider a scenario in multi-domain operations where the users in one domain need to perform searches on information hosted by a provider in another domain. It is very common in many scenarios that information cannot be shared openly across different d omains, a nd u sers m ay w ant t o o bfuscate t heir s earches a nd prevent the search provider from learning the intent of their searches. In scenarios where search privacy is important, the use of a large language model can help implement an obfuscation approach relying on generation of decoy queries to obfuscate the real query. In this paper, we consider different a lternative a pproaches t o u se l arge language models for search privacy, compare their strengths and weaknesses, and discuss their effectiveness.