A multi-layered defense against adversarial attacks in brain tumor classification using ensemble adversarial training...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Neuroscience

Background:

• Deep learning, especially Convolutional Neural Networks (CNNs), shows promise in brain tumor classification from medical images.
• However, these AI models are susceptible to adversarial attacks, which can undermine their clinical reliability.
• Vulnerability to such attacks poses a significant challenge for the safe deployment of AI in healthcare.

Purpose of the Study:

• To evaluate the robustness of a VGG16-based CNN model for brain tumor classification against adversarial attacks.
• To develop and assess a multi-layered defense strategy to enhance the model's resilience.
• To investigate the impact of defense mechanisms on AI model performance under adversarial conditions.

Main Methods:

• A VGG16 CNN model was trained for brain tumor classification on clean Magnetic Resonance Imaging (MRI) data.
• The model's performance was evaluated after exposure to Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) adversarial attacks.
• A defense strategy involving adversarial training (using FGSM/PGD examples) and feature squeezing (bit-depth reduction, Gaussian blurring) was implemented.

Main Results:

• The VGG16 model achieved 96% accuracy on clean MRI data.
• Adversarial attacks significantly reduced accuracy to 32% (FGSM) and 13% (PGD).
• The implemented defense strategy improved accuracy to 54% (FGSM) and 47% (PGD) against adversarial examples.

Conclusions:

• Deep learning models for medical image analysis, while accurate, require robust defense mechanisms against adversarial attacks.
• Proactive defense strategies are crucial for ensuring the reliability and clinical applicability of AI in medical diagnostics.
• This study demonstrates the effectiveness of combined adversarial training and feature squeezing in enhancing AI model resilience in medical imaging.