mlsanityData to 13 Oct 2026

Revisiting adversarial training for the worst-performing class

arXiv 2302.08872 ↗First posted 17 Feb 2023cs.LGcited by 4

Not on the map. Why?

Abstract (arXiv)

Despite progress in adversarial training (AT), there is a substantial gap between the top-performing and worst-performing classes in many datasets. For example, on CIFAR10, the accuracies for the best and worst classes are 74% and 23%, respectively. We argue that this gap can be reduced by explicitly optimizing for the worst-performing class, resulting in a min-max-max optimization formulation. Our method, called class focused online learning (CFOL), includes high probability convergence guarantees for the worst class loss and can be easily integrated into existing training setups with minimal computational overhead. We demonstrate an improvement to 32% in the worst class accuracy on CIFAR10, and we observe consistent behavior across CIFAR100 and STL10. Our study highlights the importance of moving beyond average accuracy, which is particularly important in safety-critical applications.

Key facts. Cited by 4 papers in the map (data through 13 Oct 2026). Builds on 28 of its 43 references that are in the map. First posted on arXiv 17 Feb 2023.

Builds on

28 of 43 references.

References · 28

  1. 2015Deep Residual Learning for Image Recognitioncited by 51,950
  2. 2014Very Deep Convolutional Networks for Large-Scale Image Recognitioncited by 15,321
  3. 2014ImageNet Large Scale Visual Recognition Challengecited by 8,476
  4. 2014Explaining and Harnessing Adversarial Examplescited by 7,336
  5. 2017Towards Deep Learning Models Resistant to Adversarial Attackscited by 6,379
  6. 2013Intriguing properties of neural networkscited by 5,137
  7. 2019Theoretically Principled Trade-off between Robustness and Accuracycited by 1,511
  8. 2020Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attackscited by 1,445
  9. 2019Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalizationcited by 1,242
  10. 2015Distillation as a Defense to Adversarial Perturbations against Deep Neural Networkscited by 676
  11. 2017Countering Adversarial Images using Input Transformationscited by 519
  12. 2019Fair Resource Allocation in Federated Learningcited by 390
16 more
  1. 2018Learning Models with Uniform Performance via Distributionally Robust Optimizationcited by 335
  2. 2018Certified Defenses against Adversarial Examplescited by 305
  3. 2020Large-Scale Methods for Distributionally Robust Optimizationcited by 179
  4. 2020Perceptual Adversarial Robustness: Defense Against Unseen Threat Modelscited by 137
  5. 2020Tilted Empirical Risk Minimizationcited by 104
  6. 2016Defensive Distillation is Not Robust to Adversarial Examplescited by 60
  7. 2017Learning with Average Top-k Losscited by 52
  8. 2019Ordered SGD: A New Stochastic Optimization Framework for Empirical Risk Minimizationcited by 51
  9. 2019Adaptive Sampling for Stochastic Risk-Averse Learningcited by 39
  10. 2018Evaluating and Understanding the Robustness of Adversarial Logit Pairingcited by 34
  11. 2020Learning Bounds for Risk-sensitive Learningcited by 33
  12. 2021Analysis and Applications of Class-wise Robustness in Adversarial Trainingcited by 33
  13. 2016Minimizing the Maximal Loss: How and Why?cited by 30
  14. 2020Class-Weighted Classification: Trade-offs and Robust Approachescited by 30
  15. 2021Reweighting Augmented Samples by Minimizing the Maximal Expected Losscited by 21
  16. 2019Is AmI (Attacks Meet Interpretability) Robust to Adversarial Examples?cited by 7

Built on by

Cited by year

Citing papers per year: 2023: 2; 2024: 0; 2025: 2.

4 papers.

  1. 2023Doubly Robust Instance-Reweighted Adversarial Trainingcited by 2
  2. 2025Towards Fair Class-wise Robustness: Class Optimal Distribution Adversarial Trainingcited by 2
  3. 2023Provably Cost-Sensitive Adversarial Defense via Randomized Smoothingcited by 0
  4. 2025Deep Minimax Classifiers for Imbalanced Datasets with a Small Number of Minority Samplescited by 0