LatAtk: A Medical Image Attack Method Focused on Lesion Areas with High Transferability.

Area of Science:

• Medical Imaging
• Deep Learning Security
• Adversarial Attacks

Background:

• Concerns regarding the security and reliability of deep learning models in sensitive healthcare applications are rising.
• Adversarial attacks pose a significant threat to the integrity of deep neural networks (DNNs), particularly in medical imaging.
• Existing attack methods lack sufficient transferability and concealment, necessitating advanced security solutions.

Purpose of the Study:

• To propose a novel adversarial attack method, LatAtk, specifically designed for medical images.
• To enhance the security and controllability of DNNs used in life and health safety applications.
• To improve the transferability and concealment of adversarial attacks against medical image analysis.

Main Methods:

• LatAtk utilizes image segmentation to differentiate between attackable (lesion) and non-attackable areas within medical images.
• Perturbations are strategically injected into lesion areas to disrupt DNN attention mechanisms.
• A class activation loss function (Gradient-Weighted Class Activation Mapping) and a texture feature loss function (Local Binary Patterns) are employed to enhance transferability and preserve image integrity.

Main Results:

• LatAtk demonstrated superior aggressiveness in disrupting DNNs compared to existing methods.
• The proposed method exhibited high transferability, effectively attacking different DNN models.
• LatAtk achieved improved concealment, preserving essential texture features and reducing detectability of adversarial samples.

Conclusions:

• LatAtk presents a significant advancement in adversarial attack methodologies for medical imaging.
• The technique offers enhanced security and reliability for DNNs in critical healthcare scenarios.
• Further research into LatAtk can drive innovation in robust defense mechanisms for medical AI.