Skip to main content
Back

Duplicate Images Silently Destroy Robust AI Models, KU Leuven Finds

New research shows duplicated images in training sets barely help standard image classifiers and actively wreck adversarially trained models on CIFAR-10.

Pro

  • Duplicates in image training sets hurt accuracy and destroy adversarial robustness, unlike the modest gains seen in LLMs.
  • Standard CIFAR-10 model: 70.72% baseline accuracy, only 75.11% at 60% duplication, then diminishing returns.
  • Adversarially trained model: accuracy collapses from 41.16% to 20.90% as duplication rises to 60%.
  • Non-uniform duplication skews decision boundaries toward over-represented classes, hurting overall generalization.
  • Theoretical bias-variance analysis explains why duplicates compound sensitivity under adversarial perturbations.
  • Code available at github.com/alireza1375/Duplication; paper at arxiv.org/abs/2504.00638.

For language models, deduplicating training data is by now standard hygiene. Removing repeated sequences boosts accuracy, reduces memorization, and makes models less vulnerable to extraction attacks. What about image classifiers? A new paper from researchers at KU Leuven, Newcastle, Essex, and Penn State takes a hard look at that question and finds the answer is surprisingly ugly: duplicates in your image training set rarely help, sometimes silently hurt, and can catastrophically degrade robust models trained against adversarial attacks.

The gap nobody was looking at

Prior work on image duplication has almost entirely focused on whether images in the training set also appear in the test set. That question matters for benchmark validity, but it dodges the practical one: what happens when your training set itself contains repeats.

The authors hypothesize that duplicates within the training set inhibit general learning and degrade efficiency, accuracy, and robustness against attacks such as adversarial perturbations. To test this they combine a theoretical bias-variance analysis, a toy setup with two Gaussian clusters classified by an SVM, and real experiments on CIFAR-10 with both standard and adversarially trained convolutional networks.

Two flavors of duplication, both bad

The study distinguishes between uniform duplication, where repeated samples are drawn randomly across all classes, and