Robin Carpentier

A global sensitivity analysis of key LLM functions, which bounds the required scale of differentially private noise and derives an upper bound on floating-point error from masking and noise cancellation in the TEE as a function of the privacy parameter epsilon.