Vulnerability Analysis of Chest X-Ray Image Classification Against Adversarial Attacks

Saeid Asgari Taghanaki, Arkadeep Das, Ghassan Hamarneh

TL;DR

This paper investigates the vulnerability of deep learning models used for chest X-ray classification against adversarial attacks. By testing ten different attack methods on two state-of-the-art networks, the authors reveal significant weaknesses, particularly with gradient-based attacks, and explore how modifying pooling operations can enhance model resilience.

What Problem Does This Paper Address?

Deep learning has shown great promise in automatically classifying chest X-ray images, aiding in the diagnosis of various lung ailments. However, the reliability and trustworthiness of these models in clinical settings are crucial. A significant concern for deep learning models in general is their susceptibility to "adversarial attacks" – subtle, often imperceptible perturbations to input data that can cause misclassification. Despite the critical nature of medical imaging, a comprehensive analysis of chest X-ray classification models' vulnerability to these attacks was lacking. This paper aims to fill that gap by systematically evaluating how well these models withstand various adversarial attacks, thereby informing the development of more robust and trustworthy AI in healthcare.

Methods

The researchers employed a two-pronged approach:

  1. Adversarial Attack Evaluation: They tested ten different adversarial attack methods, categorized into three main types:
  2. Gradient-based attacks: These attacks leverage the model's gradients to find the most sensitive parts of an image and perturb them to cause misclassification. Examples include FGSM, PGD, and DeepFool.
  3. Score-based attacks: These methods use the model's confidence scores (e.g., softmax probabilities) to estimate gradients and craft attacks. Local Search and Single Pixel attacks were used here.
  4. Decision-based attacks: These are "black-box" attacks that only rely on the model's predicted class label, without access to gradients or scores. Gaussian Blur, Contrast Reduction, and Additive Gaussian Noise were employed. Both "white-box" (attacker has full knowledge of the model) and "black-box" (attacker has limited knowledge) scenarios were considered.

  5. Pooling Operation Modification: To assess model resilience, the authors modified the pooling operations within the two chosen deep learning architectures. Specifically, they replaced the standard "max-pooling" layers with "average-pooling" layers. The hypothesis was that average-pooling, which considers a broader context of features, might be more robust to subtle perturbations than max-pooling, which focuses on the most prominent feature in a local region.

Datasets

The primary dataset used in this study is Chest X-ray14 [20]. This dataset comprises 112,120 grayscale chest X-ray images from 30,805 patients, annotated with 14 disease labels and a "no-finding" label. For their experiments, the authors treated the problem as a binary classification task ("disease" vs. "non-disease"). They randomly split the dataset into training (95,128 images), validation (16,792 images), and testing (200 images) sets. The test set was specifically curated to include both diseased and "no-finding" cases.

Key Insights and Ideas

  • Single Pixel Attack Ineffectiveness: The "single pixel attack" was completely ineffective on grayscale chest X-ray images. This suggests that while effective on RGB images, this method struggles to fool models when only a single grayscale pixel is altered, highlighting domain-specific vulnerabilities.
  • Pooling Matters: The choice of pooling operation (max vs. average) significantly impacts model resilience to certain attacks. Average-pooling generally showed improved robustness, particularly against decision and score-based attacks.
  • Gradient-Based Attacks are Potent: Gradient-based attacks proved to be the most successful in fooling both the models and, in some cases, human perception, indicating a critical vulnerability in current deep learning models for medical imaging.
  • Human Perception vs. Model Accuracy: Some adversarial attacks, like Gaussian blur and contrast reduction, create perturbations that are easily detectable by the human eye, suggesting that models can be fooled by visually obvious changes.

Key Results

  • White-Box Attacks: Gradient-based attacks were highly successful in fooling both networks (with and without pooling modification) in white-box scenarios, often reducing accuracy to near zero. Decision-based and score-based attacks were largely unsuccessful in these scenarios.
  • Black-Box Attacks: In black-box scenarios, most attacks were only partially successful, with performance drops not as severe as in white-box attacks. Average-pooling demonstrated more resilience against gradient-based black-box attacks.
  • Pooling Impact: Nasnet-Large with average-pooling was observed to be 18% stronger than its max-pooling counterpart against certain attacks. The local search attack (S1) completely failed against Nasnet-Large with average-pooling.
  • Accuracy Drop: For gradient-based black-box attacks, the authors noted that for 23%±9% and 27%±8% of test samples, both networks (with average and max pooling, respectively) failed on the same cases.

Limitations and Future Work

  • Limited Network Scope: The study focused on only two specific deep learning architectures (Inception-ResNet-v2 and Nasnet-Large). A broader evaluation across more diverse architectures would provide a more comprehensive understanding.
  • Dataset Size for Testing: While the Chest X-ray14 dataset is large, the final test set comprised only 200 images. A larger, more diverse test set could offer more robust statistical validation.
  • Human Perception Analysis: While some attacks were noted as visually detectable, a more formal human perception study would be beneficial to quantify this aspect.

The authors suggest that future work could involve developing more robust defense mechanisms and further exploring the impact of different architectural choices and training strategies on adversarial robustness in medical imaging.

Summary updated

BibTeX
@inproceedings{taghanaki2018vulnerability,
  author = {Saeid Asgari Taghanaki and Arkadeep Das and Ghassan Hamarneh},
  title = {{Vulnerability Analysis of Chest X-Ray Image Classification Against Adversarial Attacks}},
  booktitle = {MLCN/DLF/iMIMIC @ MICCAI},
  year = {2018},
  url = {https://openreview.net/forum?id=FxdaXvJ9S8},
}

Loading related content...

Figure