PAC-Private Responses with Adversarial Composition

Xiaochen Zhu, Mayuri Sridhar, Srinivas Devadas · 2026

Read PAC-Private Responses with Adversarial Composition by Xiaochen Zhu, Mayuri Sridhar, Srinivas Devadas (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-14033.txt.

Description

Xiaochen Zhu, Mayuri Sridhar, and Srinivas Devadas protect the answers of AI models served through APIs in 2026. Modern models are increasingly deployed behind such interfaces. Standard weight privatization, like DP-SGD, adds noise to the whole model. That can be unnecessarily noisy at the cost of utility.

Model responses to specific inputs are much lower dimensional and more stable. That fact motivates privacy guarantees directly on model outputs. Their PAC-Private responses handle adversarial composition, where many queries combine. Output-level privacy could keep useful detail while hiding training data. The work keeps the guard on each reply, not on the whole model at once. That focus can save much of the sharpness users value.

Catalog record

CREATORXiaochen Zhu, Mayuri Sridhar, Srinivas Devadas
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE95 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-14033

Browse the full library · About the Tower of Babel