EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

DP 2 O: Dual-Pair Direct Preference Optimization for Preference-Guided Adversarial Robustness in Vision Models

Ahmed Dawod Mohammed Ibrahum, Jang-Eui Hong

IEEE Access · 2026 · 14 · 92399-92413

DOI: 10.1109/access.2026.3701884 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Ahmed Dawod Mohammed Ibrahum und 1 weiteren Autor:innen. Die Arbeit nennt 30 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Research Foundation of Korea (NRF) grant, Ministry of Education.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Adversarial training remains the most effective empirical defense against adversarial examples, yet it is hindered by the high cost of inner maximization, the clean–robust accuracy trade-off, and the underutilization of information contained in misclassified or near-boundary adversarial samples. We propose DP2O, a preference-guided framework that converts adversarial evaluations into structured preference pairs and operationalizes Direct Preference Optimization (DPO) for vision robustness. Specifically, we are the first to import the implicit-reward, reference-anchored DPO formulation into the adversarial robustness setting for vision models, and we do so through a dual-pool construction in which adversarial failures form unstable pairs (U-pairs) preferring the ground-truth label over the model’s incorrect prediction, while adversarial successes form resilient pairs (R-pairs) preferring the ground-truth label over the runner-up class. A hybrid objective then combines DPO on these typed pairs with cross-entropy on resilient examples. Across CIFAR-10, CIFAR-100, and STL-10 using ResNet-18 and a diverse attack suite (FGSM, PGD8,20, PGD64,40, EOTPGD, Jitter, and AutoAttack), DP2O consistently improves adversarial robustness over strong baselines while maintaining competitive clean accuracy. On CIFAR-10, the hybrid model improves PGD8,20 robustness from 0.4853 (TRADES) to 0.5846, increases FGSM robustness to 0.6718, and achieves 0.4934 under AutoAttack while maintaining 0.9125 clean accuracy. Similar gains are observed on CIFAR-100 and STL-10, where DP2O improves robustness across both iterative and adaptive attacks. These results demonstrate that reframing adversarial robustness as preference learning effectively reuses adversarial outcomes as supervision and improves the clean–robust trade-off without requiring additional critic models.

Förderung

  • National Research Foundation of Korea (NRF) grant
  • Ministry of Education

Zitieren

APA

Ibrahum, A. D. M., & Hong, J. (2026). DP 2 O: Dual-Pair Direct Preference Optimization for Preference-Guided Adversarial Robustness in Vision Models. IEEE Access, 14, 92399-92413. https://doi.org/10.1109/access.2026.3701884

Vancouver

Ibrahum ADM, Hong J. DP 2 O: Dual-Pair Direct Preference Optimization for Preference-Guided Adversarial Robustness in Vision Models. IEEE Access. 2026;14:92399-92413. doi:10.1109/access.2026.3701884

BibTeX
@article{Ibrahum2026,
  author = {Ibrahum, Ahmed Dawod Mohammed and Hong, Jang-Eui},
  title = {DP 2 O: Dual-Pair Direct Preference Optimization for Preference-Guided Adversarial Robustness in Vision Models},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {92399--92413},
  doi = {10.1109/access.2026.3701884}
}
Kostenlos starten →