DEEP LEARNING FOR BRAIN TUMOR CLASSIFICATION: ROBUSTNESS AGAINST ADVERSARIAL ATTACKS AND DEFENSIVE STRATEGIES

Date

2025-12-30

Author

Khalid, Hafsa

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

179
views

0
downloads

Cite This

In this study, a Customized Convolutional Neural Network (C-CNN) and two transfer learning–based models, ResNet-50 and EfficientNetV2S, are proposed for classifying three brain tumor types - meningioma, glioma, and pituitary, using the publicly available Figshare brain tumor MRI dataset. The robustness of the models is assessed under six adversarial attacks that reflect real-world degradations in clinical MRI data, including Fast Gradient Sign Method (FGSM), motion blur, partial occlusion, JPEG compression artifacts, Gaussian noise artifacts, and adversarial boundary noise. To reduce the vulnerability of deep learning models to adversarial perturbations, two defense strategies are implemented. The first, adversarial attack-driven data augmentation, incorporates adversarially attacked images into the clean training dataset to enhance robustness. The second, defensive distillation, transfers knowledge from a teacher model to a student model by learning softened probability outputs. On the clean dataset, C-CNN, ResNet-50, and EfficientNetV2S achieved test accuracies of 94.42%, 98.03%, and 85%, respectively, using a 70% training, 15% validation, and 15% testing split. However, all models exhibited substantial performance degradation under adversarial attacks, demonstrating their limited robustness and the need for effective defense mechanisms. After applying the defense strategies, model performance was evaluated using two approaches. In the first, clean and attacked images were randomly distributed across all datasets. Under this setting, adversarial data augmentation achieved test accuracies of 97.15%, 98.30%, and 88.23% for C-CNN, ResNet-50, and EfficientNetV2S, respectively, while defensive distillation achieved 95.09%, 97.63%, and 84.95%. In the second approach, a fixed 16:84 ratio of attacked to clean images was maintained. Here, data augmentation accuracies decreased to 95.82% for C-CNN and 97.63% for ResNet-50, while EfficientNetV2S improved to 85.94%. Defensive distillation resulted in accuracies of 94.36%, 97.51%, and 84.18%, respectively.

Subject Keywords

Adversarial Attacks, Artificial Intelligence (AI), Brain tumor (BT), Customized Convolutional Neural Network (C-CNN), Deep Learning (DL)

URI

https://hdl.handle.net/11511/118756

Collections

Northern Cyprus Campus, Thesis