Auto encoder-based defense mechanism against popular adversarial attacks in deep learning

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Cybersecurity in Healthcare

Background:

• Convolutional Neural Network (CNN) models are vulnerable to adversarial attacks, where subtle image perturbations cause misclassification, posing security risks in clinical settings.
• Existing defense mechanisms often lack effectiveness against multiple attack types, necessitating a more robust solution for secure medical image analysis.
• Reliable defense frameworks are crucial for the safe clinical deployment of deep learning models, aiding diagnosis and automating tasks.

Purpose of the Study:

• To develop and evaluate a robust defense mechanism against multiple adversarial attack types for CNN-based pneumonia detection in chest X-ray images.
• To compare the performance of the proposed defense strategy against state-of-the-art attacks and defense mechanisms.
• To ensure the secure clinical deployment of deep learning models for medical image classification.

Main Methods:

• A convolutional autoencoder was employed to denoise adversarial images generated by Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks at various magnitudes (epsilon values).
• Performance was evaluated using two pre-trained models (VGG16, VGG19) and a hybrid model (Stack Model: MobileNetV2 + DenseNet169).
• The study analyzed the attack success rate and the improvement in model accuracy after applying the proposed defense mechanism.

Main Results:

• The Projected Gradient Descent (PGD) attack demonstrated a high success rate, reducing VGG16 model accuracy by up to 67%.
• The proposed convolutional autoencoder defense mechanism significantly improved accuracy, increasing it by up to 16% against PGD attacks for both VGG16 and VGG19 models.
• The defense framework demonstrated superior performance compared to existing state-of-the-art methods.

Conclusions:

• The developed defense mechanism effectively mitigates multiple adversarial attacks on CNN-based medical image analysis.
• The proposed approach enhances the reliability and robustness of pneumonia detection systems, paving the way for secure clinical integration.
• This research contributes to the secure application of artificial intelligence in healthcare, improving diagnostic accuracy and operational efficiency.