On Frequency Domain Adversarial Vulnerabilities of Volumetric Medical Image Segmentation (ISBI'25)

On Frequency Domain Adversarial Vulnerabilities of Volumetric Medical Image Segmentation

Asif Hanif, Muzammal Naseer, Salman Khan, and Fahad Shahbaz Khan

---

Spectrum Adversarial Attack $(\mathsf{SA}^2)$ partitions the clean volumetric image into 3D patches, applies 3D-DCT to each patch, and amplifies/attenuates the DCT coefficients using multiplicative spectral noise $(\xi)$. The perturbed spectrum is then converted back to the voxel domain via 3D-IDCT. The loss gradient flow to $\xi$ is shown by the black dashed line.

---

Abstract

In safety-critical domains like healthcare, resilience of deep learning models towards adversarial attacks is crucial. Volumetric medical image segmentation is a fundamental task, providing critical insights for diagnosis. This paper introduces a novel frequency domain adversarial attack targeting 3D medical data, revealing vulnerabilities in segmentation models. By manipulating the frequency spectrum (low, middle, and high bands), we assess its impact on model performance. Unlike pixel-based 2D attacks, our method continuously perturbs 3D samples with minimal information loss, achieving high fooling rates at lower computational costs and superior black-box transferability, while maintaining perceptual quality.

TLDR: A novel frequency domain attack on 3D medical segmentation exposes vulnerabilities, achieving high fooling rates, low cost, and superior transferability while preserving perceptual quality.

Updates 🚀

• Jan 03, 2025 : Accepted in ISBI 2025    🎊 🎉
• Jan 03, 2025 : Code to be released soon