Useful nonrobust features are ubiquitous in biomedical images

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Mouton, Coenraad, Rabe, Randle, Koser, Niklas C., Krekiehn, Nicolai, Hansen, Christopher, Hövener, Jan-Bernd, Glüer, Claus-C.
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911621496963072
author Mouton, Coenraad
Rabe, Randle
Koser, Niklas C.
Krekiehn, Nicolai
Hansen, Christopher
Hövener, Jan-Bernd
Glüer, Claus-C.
author_facet Mouton, Coenraad
Rabe, Randle
Koser, Niklas C.
Krekiehn, Nicolai
Hansen, Christopher
Hövener, Jan-Bernd
Glüer, Claus-C.
contents We study whether deep networks for medical imaging learn useful nonrobust features - predictive input patterns that are not human interpretable and highly susceptible to small adversarial perturbations - and how these features impact test performance. We show that models trained only on nonrobust features achieve well above chance accuracy across five MedMNIST classification tasks, confirming their predictive value in-distribution. Conversely, adversarially trained models that primarily rely on robust features sacrifice in-distribution accuracy but yield markedly better performance under controlled distribution shifts (MedMNIST-C). Overall, nonrobust features boost standard accuracy yet degrade out-of-distribution performance, revealing a practical robustness-accuracy trade-off in medical imaging classification tasks that should be tailored to the requirements of the deployment setting.
format Preprint
id arxiv_https___arxiv_org_abs_2604_22579
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Useful nonrobust features are ubiquitous in biomedical images
Mouton, Coenraad
Rabe, Randle
Koser, Niklas C.
Krekiehn, Nicolai
Hansen, Christopher
Hövener, Jan-Bernd
Glüer, Claus-C.
Image and Video Processing
Computer Vision and Pattern Recognition
Machine Learning
We study whether deep networks for medical imaging learn useful nonrobust features - predictive input patterns that are not human interpretable and highly susceptible to small adversarial perturbations - and how these features impact test performance. We show that models trained only on nonrobust features achieve well above chance accuracy across five MedMNIST classification tasks, confirming their predictive value in-distribution. Conversely, adversarially trained models that primarily rely on robust features sacrifice in-distribution accuracy but yield markedly better performance under controlled distribution shifts (MedMNIST-C). Overall, nonrobust features boost standard accuracy yet degrade out-of-distribution performance, revealing a practical robustness-accuracy trade-off in medical imaging classification tasks that should be tailored to the requirements of the deployment setting.
title Useful nonrobust features are ubiquitous in biomedical images
topic Image and Video Processing
Computer Vision and Pattern Recognition
Machine Learning
url https://arxiv.org/abs/2604.22579