Abstract
Recently, there have been several successful deep learning approaches for automatically classifying chest X-ray images into different disease categories. However, there is not yet a comprehensive vulnerability analysis of these models against the so-called adversarial perturbations/attacks, which makes deep models more trustful in clinical practices. In this paper, we extensively analyzed the performance of two state-of-the-art classification deep networks on chest X-ray images. These two networks were attacked by three different categories (ten methods in total) of adversarial methods (both white- and black-box), namely gradient-based, score-based, and decision-based attacks. Furthermore, we modified the pooling operations in the two classification networks to measure their sensitivities against different attacks, on the specific task of chest X-ray classification.
Introduction
Deep learning (DL) has demonstrated significant potential in automating the diagnosis of lung diseases from chest X-rays. However, for these models to be safely integrated into clinical workflows, they must be robust against adversarial attacks—small, often imperceptible perturbations designed to mislead the model into making incorrect predictions. While adversarial vulnerability is well-documented in general computer vision, its impact on medical imaging requires specialized investigation. This study provides a comprehensive vulnerability analysis of state-of-the-art medical AI models and explores architectural modifications to enhance their resilience.
Methodology
The researchers evaluated two prominent architectures: Inception-ResNet-v2 and Nasnet-Large. These models were subjected to ten different adversarial attacks, categorized by the level of information available to the attacker:
- Gradient-based (White-box): Attacks like FGSM and DeepFool that utilize the model's internal weights.
- Score-based (Black-box): Attacks such as "Local Search" and "Single Pixel" that rely on output confidence scores.
- Decision-based (Black-box): Attacks that only require the final predicted label, such as Gaussian noise or contrast reduction.
A key structural defense tested was the replacement of standard max-pooling layers with average-pooling. The hypothesis was that average-pooling, by considering the global context of a region rather than a single peak value, would be harder for an attacker to manipulate. The study utilized the ChestX-ray14 dataset, focusing on a binary classification task ("disease" vs. "no finding") using a test set of 200 images for the adversarial experiments.
Visualizing Adversarial Perturbations
Adversarial attacks vary significantly in their visual impact. Some are completely invisible to the human eye, while others significantly degrade image quality.
As shown in Figure 1, gradient-based attacks (G1-G5) produce perturbations that are virtually imperceptible, yet they are highly effective at fooling the model. In contrast, decision-based attacks (D1-D3) often result in obvious visual artifacts like blurring or contrast changes, which a radiologist would likely notice in a clinical setting.
Key Results and Insights
The vulnerability analysis yielded several critical findings:
- White-box Vulnerability: In white-box scenarios, gradient-based attacks were devastatingly effective, reducing model accuracy from approximately 70% to near 0% in most cases.
- The Grayscale Advantage: The "Single Pixel" attack, typically effective on color images, failed on grayscale X-rays. The limited "budget" of a single channel in grayscale images appears insufficient to reliably fool these deep models.
- Pooling for Resilience: Modifying the pooling strategy proved beneficial. Models using average-pooling (AP) demonstrated significantly higher resilience to score-based and decision-based attacks compared to those using max-pooling (MP). For instance, Nasnet-Large with AP showed a marked improvement in robustness over its MP counterpart.
Figure 4 highlights the disparity in attack effectiveness. While "Obvious" and "Mixed" attacks allow for some remaining accuracy, "Hard" gradient-based attacks consistently drive accuracy to zero, highlighting the extreme vulnerability of models when their internal gradients are exposed.
Discussion and Future Work
While architectural changes like average-pooling offer a degree of protection against black-box attacks, they do not fully mitigate the threat of white-box gradient attacks. Furthermore, the "obvious" nature of some attacks suggests that human-in-the-loop systems could catch some adversarial attempts, but the "invisible" attacks remain a significant concern for fully automated systems.
Future research should explore more robust defense mechanisms, such as adversarial training or specialized input denoising, and extend the analysis to multi-label classification tasks to better reflect real-world clinical complexity.
Conclusion
This study underscores the fragility of current medical DL models to adversarial perturbations. While simple architectural changes like average-pooling can improve robustness, a significant gap remains in defending against sophisticated white-box attacks. Ensuring the security and reliability of these models is paramount for their successful deployment in healthcare.
Summary updated