Enregistré dans:
Détails bibliographiques
Auteurs principaux: Ashiq, Muhammad H., Triantafillou, Peter, Tseng, Hung Yun, Chrysos, Grigoris G.
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2509.11625
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866915522892791808
author Ashiq, Muhammad H.
Triantafillou, Peter
Tseng, Hung Yun
Chrysos, Grigoris G.
author_facet Ashiq, Muhammad H.
Triantafillou, Peter
Tseng, Hung Yun
Chrysos, Grigoris G.
contents A key concern for AI safety remains understudied in the machine learning (ML) literature: how can we ensure users of ML models do not leverage predictions on incorrect personal data to harm others? This is particularly pertinent given the rise of open-weight models, where simply masking model outputs does not suffice to prevent adversaries from recovering harmful predictions. To address this threat, which we call *test-time privacy*, we induce maximal uncertainty on protected instances while preserving accuracy on all other instances. Our proposed algorithm uses a Pareto optimal objective that explicitly balances test-time privacy against utility. We also provide a certifiable approximation algorithm which achieves $(\varepsilon, δ)$ guarantees without convexity assumptions. We then prove a tight bound that characterizes the privacy-utility tradeoff that our algorithms incur. Empirically, our method obtains at least $>3\times$ stronger uncertainty than pretraining with marginal drops in accuracy on various image recognition benchmarks. Altogether, this framework provides a tool to guarantee additional protection to end users.
format Preprint
id arxiv_https___arxiv_org_abs_2509_11625
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Inducing Uncertainty on Open-Weight Models for Test-Time Privacy in Image Recognition
Ashiq, Muhammad H.
Triantafillou, Peter
Tseng, Hung Yun
Chrysos, Grigoris G.
Machine Learning
Artificial Intelligence
Cryptography and Security
A key concern for AI safety remains understudied in the machine learning (ML) literature: how can we ensure users of ML models do not leverage predictions on incorrect personal data to harm others? This is particularly pertinent given the rise of open-weight models, where simply masking model outputs does not suffice to prevent adversaries from recovering harmful predictions. To address this threat, which we call *test-time privacy*, we induce maximal uncertainty on protected instances while preserving accuracy on all other instances. Our proposed algorithm uses a Pareto optimal objective that explicitly balances test-time privacy against utility. We also provide a certifiable approximation algorithm which achieves $(\varepsilon, δ)$ guarantees without convexity assumptions. We then prove a tight bound that characterizes the privacy-utility tradeoff that our algorithms incur. Empirically, our method obtains at least $>3\times$ stronger uncertainty than pretraining with marginal drops in accuracy on various image recognition benchmarks. Altogether, this framework provides a tool to guarantee additional protection to end users.
title Inducing Uncertainty on Open-Weight Models for Test-Time Privacy in Image Recognition
topic Machine Learning
Artificial Intelligence
Cryptography and Security
url https://arxiv.org/abs/2509.11625