Introduction
As Deep Neural Networks (DNNs) become increasingly integrated into medical diagnostic pipelines, their security and robustness are under intense scrutiny. While traditional adversarial attack research has focused on high-frequency "noise"—perturbations that are often easily detectable by human experts—this paper, presented at ICME 2021, identifies a more insidious threat. The researchers demonstrate that DNNs can be systematically fooled by manipulating the bias field, a naturally occurring intensity inhomogeneity common in medical imaging.
The Problem: Realistic vs. Artificial Perturbations
In clinical settings, X-ray images often suffer from gradual shifts in brightness caused by improper image acquisition or spatial variations in radiation. While human radiologists are trained to ignore these smooth intensity gradients, DNNs are highly sensitive to them. Most existing adversarial attacks, such as FGSM (Fast Gradient Sign Method), add "grainy" pixel-wise noise. In a medical context, such noise is an immediate red flag. The authors address this gap by proposing a method to generate realistic, spatially smooth degradations that mimic actual hardware failures or environmental conditions.
Methodology: The AdvSBF Attack
The researchers introduced the Adversarial-Smooth Bias Field Attack (AdvSBF). This approach shifts away from standard additive noise models in three fundamental ways:
- Multiplicative Modeling: Unlike standard attacks ($X + \text{noise}$), the bias field is modeled as a multiplicative factor ($X \cdot B$), reflecting how radiation is captured in physical sensors.
- Multivariate Polynomial Smoothness: To ensure the attack remains indistinguishable from natural artifacts, the field is modeled using a multivariate polynomial rather than independent pixel optimization. This constraint forces the perturbation to be spatially smooth.
- Distortion-Aware Tuning: By incorporating Thin Plate Spline (TPS) transformations, the attack can locally tune the bias field around critical anatomical regions to maximize the probability of a misdiagnosis while maintaining overall image integrity.
Key Insights: The Smoothness Paradox and Transferability
A central finding of the study is the "Smoothness Paradox." While high-frequency noise is highly effective at targeting specific model weaknesses, it is visually obvious. Smooth bias fields are harder to "aim" but are significantly more realistic.
Furthermore, the AdvSBF attack demonstrates superior transferability. Noise-based attacks often overfit to the specific quirks of a single architecture (e.g., ResNet50) and fail when tested on others. Because AdvSBF targets more fundamental image features, an attack crafted for one model is much more likely to fool another, such as MobileNet or DenseNet121. This suggests that the vulnerability is not just a model-specific artifact but a broader weakness in how DNNs process intensity gradients.
Experimental Results
The researchers evaluated their attack using a Chest X-ray dataset containing over 5,800 images. Key metrics included:
- Realism: Using the BRISQUE metric, AdvSBF images were rated as significantly more realistic than those generated by standard noise-based baselines.
- White-Box Success: On ResNet50, the attack achieved a 38.69% success rate. While lower than the near-perfect success of noise-based methods, these images are virtually indistinguishable from real X-rays to the human eye.
- Black-Box Transferability: When an attack generated for ResNet50 was tested on MobileNet, AdvSBF maintained a 7.57% success rate, whereas traditional attacks like BIM or C&W dropped to nearly 0%.
Conclusion
The AdvSBF attack highlights a critical "hidden" vulnerability in medical AI. By exploiting naturally occurring artifacts like the bias field, adversaries can craft attacks that are both effective and visually plausible. This research underscores the need for "bias-field-robust" training methods to ensure that automated diagnostic tools are resilient against both natural image degradation and malicious manipulation.
Summary updated