Related Experiment Video
Updated: Aug 24, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
How Resilient Are Deep Learning Models in Medical Image Analysis? The Case of the Moment-Based Adversarial Attack
Theodore V Maliamanis1, Kyriakos D Apostolidis1, George A Papakostas1
1MLV Research Group, Department of Computer Science, International Hellenic University, 65404 Kavala, Greece.
Biomedicines
|October 27, 2022
Summary
This study introduces Mb-AdA, a novel adversarial attack for medical image analysis that removes critical image information rather than adding noise. This method proves effective in degrading deep learning model performance while also enhancing robustness through adversarial training.
Area of Science:
- Computer Vision
- Medical Image Analysis
- Deep Learning
Background:
- Deep neural networks (DNNs) are widely used in computer vision (CV).
- Medical Image Analysis (MIA) faces significant challenges, including vulnerability to adversarial attacks (AdAs).
- Adversarial attacks threaten vision systems by degrading model performance.
Purpose of the Study:
- Propose a new black-box adversarial attack, Mb-AdA, for MIA.
- Investigate a defensive adversarial training method using Mb-AdA examples.
- Evaluate the attack's effectiveness and the defense's impact on model robustness.
Main Methods:
- Developed Mb-AdA, an attack based on orthogonal image moments.
- Applied Mb-AdA to classification and segmentation tasks on X-ray, histopathology, and cell images.
- Utilized six state-of-the-art Deep Learning models for evaluation.
Main Results:
- Mb-AdA degrades model performance up to 65% in accuracy and 18% in IoU.
- The attack preserves image structure by removing information, unlike noise-based attacks.
- Adversarial training with Mb-AdA examples enhanced model robustness.
Conclusions:
- Mb-AdA is an effective adversarial attack for medical image analysis.
- The proposed attack offers a unique approach by removing critical image information.
- Adversarial training using Mb-AdA demonstrates potential for improving model resilience.
