Title: Hierarchical Self-Supervised Adversarial Training for Robust Vision Models in Histopathology

Original: arXiv:2503.10629v2 Announce Type: replace Abstract: Adversarial attacks pose significant challenges for vision models in critical fields like healthcare, where reliability is essential. Although adversarial training has been well studied in natural images, its application to biomedical and microscopy data remains limited. Existing self-supervised adversarial training methods overlook the hierarchical structure of histopathology images, where patient-slide-patch relationships provide valuable discriminative signals. To address this, we propose Hierarchical Self-Supervised Adversarial Training (HSAT), which exploits these properties to craft adversarial examples using multi-level contrastive learning and integrate it into adversarial training for enhanced robustness. We evaluate HSAT on multiclass histopathology dataset OpenSRH and the results show that HSAT outperforms existing methods from both biomedical and natural image domains. HSAT enhances robustness, achieving an average gain of 54.31% in the white-box setting and reducing performance drops to 3-4% in the black-box setting, compared to 25-30% for the baseline. These results set a new benchmark for adversarial training in this domain, paving the way for more robust models. Our Code for training and evaluation is available at https://github.com/HashmatShadab/HSAT.

Rewrite: Adversarial attacks present a major hurdle for vision models in high-stakes sectors such as healthcare, where system dependability is paramount. While adversarial training is extensively researched for natural images, its utilization in biomedical and microscopic data has been scarce. Current self-supervised adversarial training approaches fail to account for the hierarchical organization inherent in histopathology images, ignoring the discriminative potential of patient-slide-patch relationships. To bridge this gap, we introduce Hierarchical Self-Supervised Adversarial Training (HSAT). This method leverages these structural properties to generate adversarial examples via multi-level contrastive learning, subsequently integrating them into the adversarial training framework to boost robustness. Our evaluation of HSAT on the multiclass OpenSRH histopathology dataset demonstrates that it surpasses current methods from both the biomedical and natural image fields. HSAT significantly improves robustness, yielding an average improvement of 54.31% in white-box scenarios and limiting performance degradation to just 3-4% in black-box settings, a stark contrast to the 25-30% drop observed in baseline models. These findings establish a new standard for adversarial training within this specific domain, facilitating the development of more resilient models. The code for both training and evaluation can be accessed at https://github.com/HashmatShadab/HSAT.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC