How to beat a Bayesian adversary

Deep neural networks and other modern machine learning models are often susceptible to adversarial attacks. Indeed, an adversary may often be able to change a model’s prediction through a small, directed perturbation of the model’s input – an issue in safety-critical applications. Adversarially robu...

Full description

Saved in:
Bibliographic Details
Main Authors: Zihan Ding, Kexin Jin, Jonas Latz, Chenguang Liu
Format: Article
Language:English
Published: Cambridge University Press
Series:European Journal of Applied Mathematics
Subjects:
Online Access:https://www.cambridge.org/core/product/identifier/S0956792525000105/type/journal_article
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849338212512694272
author Zihan Ding
Kexin Jin
Jonas Latz
Chenguang Liu
author_facet Zihan Ding
Kexin Jin
Jonas Latz
Chenguang Liu
author_sort Zihan Ding
collection DOAJ
description Deep neural networks and other modern machine learning models are often susceptible to adversarial attacks. Indeed, an adversary may often be able to change a model’s prediction through a small, directed perturbation of the model’s input – an issue in safety-critical applications. Adversarially robust machine learning is usually based on a minmax optimisation problem that minimises the machine learning loss under maximisation-based adversarial attacks. In this work, we study adversaries that determine their attack using a Bayesian statistical approach rather than maximisation. The resulting Bayesian adversarial robustness problem is a relaxation of the usual minmax problem. To solve this problem, we propose Abram – a continuous-time particle system that shall approximate the gradient flow corresponding to the underlying learning problem. We show that Abram approximates a McKean–Vlasov process and justify the use of Abram by giving assumptions under which the McKean–Vlasov process finds the minimiser of the Bayesian adversarial robustness problem. We discuss two ways to discretise Abram and show its suitability in benchmark adversarial deep learning experiments.
format Article
id doaj-art-907d0deeb6044426afc061c085cb7c6d
institution Kabale University
issn 0956-7925
1469-4425
language English
publisher Cambridge University Press
record_format Article
series European Journal of Applied Mathematics
spelling doaj-art-907d0deeb6044426afc061c085cb7c6d2025-08-20T03:44:28ZengCambridge University PressEuropean Journal of Applied Mathematics0956-79251469-442512310.1017/S0956792525000105How to beat a Bayesian adversaryZihan Ding0Kexin Jin1Jonas Latz2https://orcid.org/0000-0002-4600-0247Chenguang Liu3Department of Electrical and Computer Engineering, Princeton University, Princeton, USADepartment of Mathematics, Princeton University, Princeton, USADepartment of Mathematics, University of Manchester, Manchester, UKDelft Institute of Applied Mathematics, Technische Universiteit Delft, Delft, The NetherlandsDeep neural networks and other modern machine learning models are often susceptible to adversarial attacks. Indeed, an adversary may often be able to change a model’s prediction through a small, directed perturbation of the model’s input – an issue in safety-critical applications. Adversarially robust machine learning is usually based on a minmax optimisation problem that minimises the machine learning loss under maximisation-based adversarial attacks. In this work, we study adversaries that determine their attack using a Bayesian statistical approach rather than maximisation. The resulting Bayesian adversarial robustness problem is a relaxation of the usual minmax problem. To solve this problem, we propose Abram – a continuous-time particle system that shall approximate the gradient flow corresponding to the underlying learning problem. We show that Abram approximates a McKean–Vlasov process and justify the use of Abram by giving assumptions under which the McKean–Vlasov process finds the minimiser of the Bayesian adversarial robustness problem. We discuss two ways to discretise Abram and show its suitability in benchmark adversarial deep learning experiments.https://www.cambridge.org/core/product/identifier/S0956792525000105/type/journal_articleMachine learningadversarial robustnessStochastic differential equationsMcKean–Vlasov processparticle system90C1568T0765C35
spellingShingle Zihan Ding
Kexin Jin
Jonas Latz
Chenguang Liu
How to beat a Bayesian adversary
European Journal of Applied Mathematics
Machine learning
adversarial robustness
Stochastic differential equations
McKean–Vlasov process
particle system
90C15
68T07
65C35
title How to beat a Bayesian adversary
title_full How to beat a Bayesian adversary
title_fullStr How to beat a Bayesian adversary
title_full_unstemmed How to beat a Bayesian adversary
title_short How to beat a Bayesian adversary
title_sort how to beat a bayesian adversary
topic Machine learning
adversarial robustness
Stochastic differential equations
McKean–Vlasov process
particle system
90C15
68T07
65C35
url https://www.cambridge.org/core/product/identifier/S0956792525000105/type/journal_article
work_keys_str_mv AT zihanding howtobeatabayesianadversary
AT kexinjin howtobeatabayesianadversary
AT jonaslatz howtobeatabayesianadversary
AT chenguangliu howtobeatabayesianadversary