Attack-Dependent Composition Effects and Defense Selection in Adversarial Image Classification

Input-preprocessing defenses are often combined, and studies of such combinations ask whether a composition improves robustness. When the effect of adding a component changes sign across attacks, this leaves open which pipeline should be deployed. We treat the choice among preprocessing pipelines as a decision problem over a finite table of recovery rates and characterize its geometry: a pipeline is optimal for some attack weights exactly when no mixture of the other pipelines strictly dominates it, each pipeline’s winning region is a convex polytope, and preferences switch on hyperplanes. We apply this framework to Gaussian perturbation, feature squeezing, spatial smoothing, and their pairwise compositions, using DenseNet169 and ResNet-50 on ImageNetV2 (614 and 569 correctly classified images from as many classes) and four attacks generated against the undefended classifiers. On both architectures, adding Gaussian noise after median smoothing raises recovery under PGD, FGSM, and BIM by 7.0–16.4 percentage points but lowers it under Boundary Attack by 22.1–27.6 points relative to the better constituent. Mean and maximin recovery select this composition, whereas minimax regret selects smoothing alone. On both architectures, removing dominated pipelines discards the identity and feature squeezing with smoothing; for DenseNet169, feature squeezing with Gaussian perturbation wins no individual attack yet is the unique optimum for mixtures of FGSM and Boundary Attack, and it remains optimal for some attack weights in 96% of bootstrap replicates. When clean accuracy is weighted by an attack probability, no preprocessing is preferred below a probability of about 0.26 for DenseNet169 and 0.42 for ResNet-50. An adaptive BPDA/EOT attack reduces the recovery of every tested pipeline to zero, so the empirical results describe a non-adaptive threat model.

Authors

Institutions

Publication Details

Journal
Mathematics
Published
2026-11-16
DOI
https://doi.org/10.3390/math14224213
Primary Topic
Adversarial Robustness in Machine Learning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Attack-Dependent Composition Effects and Defense Selection in Adversarial Image Classification

Nataliia Kussul, Bohdan Omelchenko, Andrii Shelestov
1 links
Mathematics
Adversarial Robustness in Machine Learning
article

Attack-Dependent Composition Effects and Defense Selection in Adversarial Image Classification

Nataliia Kussul, Bohdan Omelchenko, Andrii Shelestov
article en

Abstract

Input-preprocessing defenses are often combined, and studies of such combinations ask whether a composition improves robustness. When the effect of adding a component changes sign across attacks, this leaves open which pipeline should be deployed. We treat the choice among preprocessing pipelines as a decision problem over a finite table of recovery rates and characterize its geometry: a pipeline is optimal for some attack weights exactly when no mixture of the other pipelines strictly dominates it, each pipeline’s winning region is a convex polytope, and preferences switch on hyperplanes. We apply this framework to Gaussian perturbation, feature squeezing, spatial smoothing, and their pairwise compositions, using DenseNet169 and ResNet-50 on ImageNetV2 (614 and 569 correctly classified images from as many classes) and four attacks generated against the undefended classifiers. On both architectures, adding Gaussian noise after median smoothing raises recovery under PGD, FGSM, and BIM by 7.0–16.4 percentage points but lowers it under Boundary Attack by 22.1–27.6 points relative to the better constituent. Mean and maximin recovery select this composition, whereas minimax regret selects smoothing alone. On both architectures, removing dominated pipelines discards the identity and feature squeezing with smoothing; for DenseNet169, feature squeezing with Gaussian perturbation wins no individual attack yet is the unique optimum for mixtures of FGSM and Boundary Attack, and it remains optimal for some attack weights in 96% of bootstrap replicates. When clean accuracy is weighted by an attack probability, no preprocessing is preferred below a probability of about 0.26 for DenseNet169 and 0.42 for ResNet-50. An adaptive BPDA/EOT attack reduces the recovery of every tested pipeline to zero, so the empirical results describe a non-adaptive threat model.

MathematicsVol. 14(22)
State Space Agency of Ukraine (UA), National Academy of Sciences of Ukraine (UA), National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute” (UA), Space Research Institute (UA), University of Maryland, College Park (US)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon