The context is disturbed locally, then read in pieces with compact memory. The approach seeks to limit the loss of utility due to differential confidentiality.
DP-MIND targets requests addressed to a remote model when the documents contain sensitive information. The framework applies a local disturbance and keeps a memory of the useful elements during the successive processing of the blocks.
The construction of this memory is trained on already disturbed inputs. A local module helps to restore degraded response fragments without transmitting the original context to the external service. The authors propose a theoretical analysis of confidentiality and long-context evaluations; the guarantees depend on the mechanism and the parameters chosen.
Preprint filed on arXiv on October 9, 2026; results are those reported by the authors and remain to be independently confirmed.
Read the full article on the original source
Go to the source →