Longitudinal Mammogram Exam-Based Breast Cancer Diagnosis Models: Vulnerability to Adversarial Attacks

Zhengbo Zhou, Degan Hao, Dooman Arefan, Margarita L. Zuley, Jules H. Sumkin, Lars Grimm, James Joshi, Shandong WuDeep-Breath @ MICCAI 2025 · DOI

TL;DR

This paper investigates the vulnerability of longitudinal breast cancer diagnosis models to adversarial attacks. The authors propose a novel attack method that leverages the relationship between sequential mammograms, demonstrating its effectiveness even against models trained with adversarial defenses, highlighting a critical security concern for AI in healthcare.

What Problem Does This Paper Address?

Mammography-based AI models are increasingly used for breast cancer detection and diagnosis. These models often benefit from analyzing sequential imaging exams (longitudinal data) to capture subtle changes over time, leading to improved accuracy compared to single-exam models. However, the cybersecurity of these advanced longitudinal models remains largely unexplored. Adversarial attacks, which involve subtly manipulating input data to cause misclassification, pose a significant threat to AI integrity, especially in critical applications like healthcare where misdiagnosis can have severe consequences. This paper addresses the gap in understanding how longitudinal AI models, which rely on temporal relationships, are susceptible to such attacks and proposes a method to exploit this vulnerability.

Methods

The authors propose a novel adversarial attack method designed specifically for longitudinal mammogram diagnosis models. Their approach involves crafting adversarial samples that fool a "Target" model, which uses two sequential mammogram exams (Prior and Current), by attacking it using a "Source" model trained on a single mammogram exam (Current). This "attack transferring" strategy is employed because single-exam models are more accessible to attackers.

The core of their attack lies in a two-pronged loss function:

  • Cross-entropy loss: This standard component guides the adversarial samples to cross the model's decision boundary.
  • Distance metric learning: This novel component leverages the understanding that longitudinal models use the relationship between Prior (always normal) and Current (potentially cancerous) exams. The authors exploit the fact that the Euclidean distance between Prior and Current exams is typically smaller for normal cases and larger for cancer cases. Their "knowledge-guided selection" method uses this distance to strategically select adversarial samples that are either very close to the Prior image (for cancer cases, aiming to make them appear normal) or very far from it (for normal cases, aiming to make them appear cancerous).

The attack is implemented using an iterative approach (I-FGSM) and is evaluated in a black-box setting, meaning the attacker does not have access to the Target model's architecture or parameters.

Datasets

The study was conducted on a dataset of 590 breast cancer patients, comprising 293 cancer cases and 297 cancer-free controls. Each subject provided two sequential mammogram exams, taken approximately one year apart. The Prior exams were all normal/healthy, while the Current exams contained the diagnosis outcomes, which were biopsy-proven.

Key Insights and Ideas

  • Longitudinal models are vulnerable: Contrary to the assumption that more complex models might be inherently more robust, this study demonstrates that longitudinal models, which leverage temporal information, are indeed susceptible to adversarial attacks.
  • Exploiting temporal relationships is key: The proposed attack's effectiveness stems from its ability to exploit the specific way longitudinal models use the relationship between sequential exams, rather than just attacking individual images in isolation.
  • Knowledge-guided attacks are potent: By incorporating domain knowledge about how normal and cancerous progression affects the temporal distance between mammograms, the attack can be significantly more effective.
  • Adversarial training is not a silver bullet: While adversarial training improves robustness, it doesn't fully negate the proposed attack's impact, indicating a need for more sophisticated defense mechanisms.
  • Attack transferability is practical: The ability to attack a complex longitudinal model using a simpler, more accessible single-exam model makes this research highly relevant to real-world security scenarios.

Key Results

  • The proposed "knowledge-guided" adversarial attack method significantly degraded the performance of the longitudinal breast cancer diagnosis model.
  • In the absence of adversarial training, the proposed attack achieved the lowest Area Under the ROC Curve (AUC) of 0.205, effectively fooling the model into giving opposite diagnoses (AUC < 0.5). This outperformed several state-of-the-art adversarial attacks like FGSM, I-FGSM, C&W, MI-FGSM, and PGD.
  • Even when the Target model was trained with adversarial training, the proposed attack still achieved a competitive AUC of 0.577, demonstrating its resilience against common defense strategies.
  • The study also showed that the attack remained effective across different backbone architectures (Swin Transformer and VGG) for the Source model.

Limitations and Future Work

  • Limited to specific model architecture: The study focused on a Transformer-based architecture for the Target model. Future work could explore the vulnerability of other longitudinal model architectures.
  • Focus on specific attack types: While the proposed method is novel, further research could investigate other attack strategies and their effectiveness against longitudinal models.
  • Developing robust defenses: The paper highlights the need for more effective defense methods against these sophisticated attacks. Future work should focus on developing and evaluating new adversarial defense strategies tailored for longitudinal medical imaging models.
  • Real-world deployment considerations: While the study demonstrates vulnerability, further research is needed to understand the practical implications and potential mitigation strategies for deploying these models in clinical settings.

Summary updated

BibTeX
@inproceedings{zhou2025longitudinal,
  author = {Zhengbo Zhou and Degan Hao and Dooman Arefan and Margarita L. Zuley and Jules H. Sumkin and Lars Grimm and James Joshi and Shandong Wu},
  title = {{Longitudinal Mammogram Exam-Based Breast Cancer Diagnosis Models: Vulnerability to Adversarial Attacks}},
  booktitle = {Deep-Breath @ MICCAI},
  year = {2025},
  doi = {10.1007/978-3-032-05559-0\_35},
}
Figure