mlsanityData to 14 Oct 2026

Unveiling the Backdoor Mechanism Hidden Behind Catastrophic Overfitting in Fast Adversarial Training

arXiv 2604.24350 ↗First posted 27 Apr 2026cs.LG cs.AI cs.CRcited by 0

Not on the map. Why?

Abstract (arXiv)

Fast Adversarial Training (FAT) has attracted significant attention due to its efficiency in enhancing neural network robustness against adversarial attacks. However, FAT is prone to catastrophic overfitting (CO), wherein models overfit to the specific attack used during training and fail to generalize to others. While existing methods introduce diverse hypotheses and propose various strategies to mitigate CO, a systematic and intuitive explanation of CO remains absent. In this work, we innovatively interpret CO through the lens of backdoor. Through validations on pathway division, diverse feature predictions, and universal class distinguishable triggers in CO, we conceptualize CO as a weak trigger variant of unlearnable tasks, unifying CO, backdoor attacks, and unlearnable tasks under a common theoretical framework. Guided by this, we leverage several backdoor inspired strategies to mitigate CO: (i) Recalibrate CO affected model parameters using vanilla fine tuning, linear probing, or reinitialization-based techniques; (ii) Introduce a weight outlier suppression constraint to regulate abnormal deviations in model weights. Extensive experiments support our interpretation of CO and show the efficacy of the proposed mitigation strategies.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 74 of its 98 references that are in the map. First posted on arXiv 27 Apr 2026.

Builds on

74 of 98 references.

References · 74

  1. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  2. 2014Explaining and Harnessing Adversarial Examplescited by 7,364
  3. 2017Towards Deep Learning Models Resistant to Adversarial Attackscited by 6,415
  4. 2016Towards Evaluating the Robustness of Neural Networkscited by 2,998
  5. 2018UMAP: Uniform Manifold Approximation and Projection for Dimension Reductioncited by 2,898
  6. 2016Identity Mappings in Deep Residual Networkscited by 1,965
  7. 2020Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attackscited by 1,451
  8. 2017Targeted Backdoor Attacks on Deep Learning Systems Using Data Poisoningcited by 1,055
  9. 2016Adversarial Machine Learning at Scalecited by 936
  10. 2020Fast is better than free: Revisiting adversarial trainingcited by 727
  11. 2015Cyclical Learning Rates for Training Neural Networkscited by 629
  12. 2018Fine-Pruning: Defending Against Backdooring Attacks on Deep Neural Networkscited by 580
38 more
  1. 2020Overfitting in adversarially robust deep learningcited by 506
  2. 2020Invisible Backdoor Attack with Sample-Specific Triggerscited by 353
  3. 2020Weight Poisoning Attacks on Pre-trained Modelscited by 341
  4. 2021Anti-Backdoor Learning: Training Clean Models on Poisoned Datacited by 263
  5. 2021Adversarial Neuron Pruning Purifies Backdoored Deep Modelscited by 235
  6. 2021Unlearnable Examples: Making Personal Data Unexploitablecited by 193
  7. 2024Defensive Unlearning with Adversarial Training for Robust Concept Erasure in Diffusion Modelscited by 168
  8. 2021Rethinking the Backdoor Attacks' Triggers: A Frequency Perspectivecited by 161
  9. 2024Watch Out for Your Agents! Investigating Backdoor Threats to LLM-Based Agentscited by 137
  10. 2023BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learningcited by 115
  11. 2021Backdoor Attacks on Pre-trained Models by Layerwise Weight Poisoningcited by 113
  12. 2022How to Backdoor Diffusion Models?cited by 104
  13. 2022Shadows can be Dangerous: Stealthy and Effective Physical-world Adversarial Attack by Natural Phenomenoncited by 101
  14. 2022LAS-AT: Adversarial Training with Learnable Attack Strategycited by 95
  15. 2024Enhancing Noise Robustness of Retrieval-Augmented Language Models with Adaptive Adversarial Trainingcited by 89
  16. 2024Defending Against Unforeseen Failure Modes with Latent Adversarial Trainingcited by 88
  17. 2020Understanding Catastrophic Overfitting in Single-step Adversarial Trainingcited by 76
  18. 2021Revisiting and Advancing Fast Adversarial Training Through The Lens of Bi-Level Optimizationcited by 76
  19. 2022SegPGD: An Effective and Efficient Adversarial Attack for Evaluating and Boosting Segmentation Robustnesscited by 74
  20. 2022When Adversarial Training Meets Vision Transformers: Recipes from Training to Architecturecited by 64
  21. 2023Reconstructive Neuron Pruning for Backdoor Defensecited by 59
  22. 2022AdvDO: Realistic Adversarial Attacks for Trajectory Predictioncited by 49
  23. 2022Make Some Noise: Reliable and Efficient Single-Step Adversarial Trainingcited by 43
  24. 2022Transferable Unlearnable Examplescited by 43
  25. 2023Backdoor Defense via Adaptively Splitting Poisoned Datasetcited by 43
  26. 2022Enhancing Adversarial Training with Second-Order Statistics of Weightscited by 42
  27. 2021Subspace Adversarial Trainingcited by 41
  28. 2021Boosting Fast Adversarial Training with Learnable Adversarial Initializationcited by 40
  29. 2022Unlearnable Clusters: Towards Label-agnostic Unlearnable Examplescited by 36
  30. 2022Robust Unlearnable Examples: Protecting Data Against Adversarial Learningcited by 35
  31. 2022Prior-Guided Adversarial Initialization for Fast Adversarial Trainingcited by 35
  32. 2023Improving Fast Adversarial Training with Prior-Guided Knowledgecited by 35
  33. 2024Revisiting Adversarial Training at Scalecited by 33
  34. 2023Towards Stable Backdoor Purification through Feature Shift Tuningcited by 32
  35. 2022An Embarrassingly Simple Backdoor Attack on Self-supervised Learningcited by 31
  36. 2022Towards Efficient Adversarial Training on Vision Transformerscited by 30
  37. 2023Learning the Unlearnable: Adversarial Augmentations Suppress Unlearnable Example Attackscited by 30
  38. 2021Reliably fast adversarial training via latent adversarial perturbationcited by 29

50 most cited of 74 All 74 (JSON)

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Rethinking Backdoor Adversarial Unlearning through the Lens of Catastrophic Forgetting in Continual Learning
  2. 2025Backdoor Unlearning by Linear Task Decomposition
  3. 2024Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness
  4. 2022Towards A Critical Evaluation of Robustness for Deep Learning Backdoor Countermeasures
  5. 2023Post-Training Overfitting Mitigation in DNN Classifiers
  6. 2024Backdoor in Seconds: Unlocking Vulnerabilities in Large Pre-trained Models via Model Editing
  7. 2026No Free Efficiency: Revisiting the Trade-off Between Training Efficiency and Model Vulnerability
  8. 2026Low-ASR Backdoors: Exploiting Attack Success Rate Reduction and Attacker-Defender Asymmetry ↗