A Hierarchical Feature Constraint to Camouflage Medical Adversarial Attacks

Qingsong Yao, Zecheng He, Yi Lin, Kai Ma, Yefeng Zheng, S. Kevin Zhou

TL;DR

Researchers have developed a "Hierarchical Feature Constraint" (HFC) that allows adversarial attacks on medical AI to bypass state-of-the-art detectors. By modeling normal feature distributions and forcing adversarial examples to stay within those bounds, they demonstrate that medical AI is even more vulnerable than previously thought—not just to the attacks themselves, but to attacks that are virtually invisible to statistical monitoring.

What Problem Does This Paper Address?

Deep Neural Networks (DNNs) are increasingly used for medical diagnoses, such as identifying pneumonia in X-rays or diabetic retinopathy in fundus images. However, these models are notoriously vulnerable to adversarial examples (AEs)—images with tiny, human-imperceptible perturbations that trick the AI into making a wrong call.

While these perturbations are invisible to the eye, they usually leave a "statistical footprint" in the network's internal layers. Previous research found that medical AEs are actually much easier to detect than those in natural images (like photos of cats or cars) because they tend to push the network's internal representations into "outlier" regions. This paper asks two critical questions:

  1. Why are medical AEs so much easier to detect?
  2. Can a sophisticated attacker exploit this vulnerability to "camouflage" their attack and bypass security filters?

Methods

The authors first perform a "stress test" to compare the robustness of medical vs. natural image features. They then introduce the Hierarchical Feature Constraint (HFC), a novel "plug-in" for existing white-box attacks (like BIM, PGD, or CW).

The HFC works in three main steps:

  1. Modeling Normalcy: The researchers use a Gaussian Mixture Model (GMM) to map out the "normal" distribution of features for each layer of a healthy, trained medical AI.
  2. The Camouflage Loss: They add a new mathematical constraint to the attack's objective function. Instead of just trying to flip the diagnosis (e.g., from "Pneumonia" to "Normal"), the attack now also tries to maximize the likelihood that the resulting internal features look like they belong to the "normal" distribution.
  3. Hierarchical Enforcement: Unlike previous methods that only look at the final layer, HFC applies this constraint across multiple internal layers of the network simultaneously.

Datasets

The study evaluated the HFC on two prominent public medical imaging benchmarks:

  • Kaggle Fundoscopy Dataset: 3,663 images for Diabetic Retinopathy (DR) classification.
  • Kaggle Chest X-Ray Dataset: 5,863 images for Pneumonia classification.
  • Models: The attacks were tested against standard architectures, specifically ResNet-50 and VGG-16, both of which achieved high baseline accuracy (>96%) on these tasks.

Key Insights and Ideas

  • The "Double-Edged Sword" of Vulnerability: The authors discovered that medical image features are significantly more "brittle" than natural ones. A tiny change in pixels causes a massive shift in the internal feature space. While this makes traditional attacks easy to spot, it also gives a clever attacker more "room" to move features around and hide them within the normal distribution without needing to change the image pixels significantly.
  • Gradient Consistency: The paper theoretically proves that in binary medical diagnosis tasks, the gradients used to create attacks point in a very consistent direction. This "straight line" to the outlier region is exactly what makes traditional attacks so easy for detectors to catch.
  • Hiding in the Crowd: The core "aha" moment is that an attacker doesn't need to match a specific "guide" image. They just need to ensure the adversarial features land anywhere within the high-density regions of the normal data distribution.

Key Results

The HFC proved devastatingly effective at bypassing six different state-of-the-art adversarial detectors (including KD, MAHA, and LID).

  • Bypassing Detectors: On the Fundoscopy dataset using a ResNet-50, the MAHA detector (which usually catches nearly 100% of attacks) saw its detection success (AUC) plummet from 99.6% down to 6.4% when HFC was used.
  • Superiority over Competitors: HFC significantly outperformed other "adaptive" attacks. For instance, on Chest X-Rays, HFC reduced the MAHA detector's AUC to 0.0%, whereas the next best adaptive attack only reduced it to 38.3%.
  • Semi-White-Box Success: Even when the attacker didn't know the exact parameters of the target model (only the architecture), HFC still successfully camouflaged the attacks, suggesting a high degree of transferability.

Limitations and Future Work

The primary limitation is that HFC is currently a white-box attack, meaning the attacker needs access to the model's architecture and weights to train the GMMs and calculate gradients. While the authors showed some success in "semi-white-box" scenarios, a truly "black-box" version (where the attacker has no model access) remains a challenge.

The authors suggest that future work should look into more complex ways of modeling feature distributions beyond GMMs and explore how these camouflaged attacks might affect other medical AI tasks like organ segmentation or landmark detection. This research serves as a sobering reminder that as medical AI defenses get better, the attacks are evolving to be even more deceptive.

Summary updated

Abstract

Deep neural networks for medical images are extremely vulnerable to adversarial examples (AEs), which poses security concerns on clinical decision-making. Recent findings have shown that existing medical AEs are easy to detect in feature space. To better understand this phenomenon, we thoroughly investigate the characteristic of traditional medical AEs in feature space. Specifically, we first perform a stress test to reveal the vulnerability of medical images and compare them to natural images. Then, we theoretically prove that the existing adversarial attacks manipulate the prediction by continuously optimizing the vulnerable representations in a fixed direction, leading to outlier representations in feature space. Interestingly, we find this vulnerability is a double-edged sword that can be exploited to help hide AEs in the feature space. We propose a novel hierarchical feature constraint (HFC) as an add-on to existing white-box attacks, which encourages hiding the adversarial representation in the normal feature distribution. We evaluate the proposed method on two public medical image datasets, namely Fundoscopy and Chest X-Ray. Experimental results demonstrate the superiority of our HFC as it bypasses an array of state-of-the-art adversarial medical AEs detector more efficiently than competing adaptive attacks. Our code is available at https://github.com/qsyao/Hierarchical_Feature_Constraint .
BibTeX
@inproceedings{yao2021hierarchical,
  author = {Qingsong Yao and Zecheng He and Yi Lin and Kai Ma and Yefeng Zheng and S. Kevin Zhou},
  title = {{A Hierarchical Feature Constraint to Camouflage Medical Adversarial Attacks}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  doi = {10.1007/978-3-030-87199-4\_4},
}
Figure