Reversing skin cancer adversarial examples by multiscale diffusive and denoising aggregation mechanism

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning models are crucial for skin cancer diagnosis.
• These models are vulnerable to adversarial attacks, which degrade performance.
• Adversarial attacks involve subtle image perturbations that are often imperceptible.

Purpose of the Study:

• To develop a defense framework against adversarial attacks on skin cancer diagnosis models.
• To enhance the reliability and robustness of computer-aided diagnostic systems.
• To mitigate the impact of adversarial perturbations on model performance.

Main Methods:

• A multiscale image pyramid preserves discriminative structures in medical images.
• Progressive diffusion with isotropic Gaussian noise neutralizes adversarial effects.
• A novel multiscale denoising mechanism aggregates information from neighboring scales to reverse noise.

Main Results:

• The proposed method successfully reverses adversarial perturbations across various attacks.
• Evaluated on the ISIC 2019 dataset, the framework significantly outperforms state-of-the-art defense methods.
• The defense framework is resource-efficient and effective in preserving diagnostic accuracy.

Conclusions:

• The developed defense framework offers a simple, effective solution against adversarial attacks in skin cancer diagnosis.
• This approach enhances the reliability of deep learning models for medical imaging applications.
• The method shows promise for improving the security and trustworthiness of AI-driven diagnostic tools.